Innovation, Emissions Policy, and Competitive Advantage

in the Diffusion of European Diesel Automobiles ?

Eugenio J. Miravete† Marıa J. Moral‡ Jeff Thurk§

May 2016

Abstract

Volkswagen’s recent admission that it knowingly deceived consumers and EPA regulators through

a sophisticated scheme to mask emissions from its diesel vehicles resulted in significant financial

penalties for the firm. We use this episode as motivation to assess the degree to which vehicle

emissions policies impact the automobile industry, focusing on the European market where

diesels are popular. Using Spanish automobile registration data, we estimate a discrete choice,

oligopoly model of horizontally differentiated products. Our estimation uses changes in observed

product characteristics to identify the underlying demand and cost parameters while allowing

for correlation between observed and unobserved (to the researcher) product characteristics.

We find the EU emissions policy promoted diesel vehicles by setting weaker thresholds for the

emissions produced by these vehicles. Further, diesels amounted to an important competitive

advantage for European auto makers over foreign imports. We conclude that the emissions

policy employed by EU regulators amounted to a significant non-tariff trade policy equivalent

to a 13 to 16% import tariff. Imposing product characteristic exogeneity in the estimation leads

the researcher to over-state these effects.

Keywords: Trade Policy, Import Tariff Equivalence, Diesel Cars, Emission Standards.

JEL Codes: O33, L62, F13.

?

We thank Allan Collard-Wexler, Kenneth Hendricks, Tom Holmes, Cristian Huse, Amil Petrin, Jeff Prince, andaudiences at several seminars and conferences. We are solely responsible for any errors that may still remain. Moralgratefully acknowledges funding from the Spanish Ministry of Education and Science through grant ECO2014-52051-R and ECO2015-69334.

† The University of Texas at Austin, Department of Economics, 2225 Speedway Stop 3100, Austin, Texas 78712-0301;Centre for Competition Policy/UEA; and CEPR, London, UK. Phone: 512-232-1718. Fax: 512-471-3510. E–mail:miravete@eco.utexas.edu; http://www.eugeniomiravete.com

‡ Facultad de Ciencias Economicas y Empresariales, UNED, Paseo Senda del Rey, 11, 28040, Madrid, Spain;and GRiEE, Vigo, Spain; Phone: +34-91-398-8930. Fax: +34-91-398-7821. E–mail: mjmoral@cee.uned.es;http://webs.uvigo.es/mjmoral

§ University of Notre Dame, Department of Economics, 919 Flanner Hall, Notre Dame, IN 46556. Phone: 574-631-3083. E–mail: jthurk@nd.edu; http://www.nd.edu/ jthurk/

1 Introduction

On September 18th, 2015 the United States Environmental Protection Agency (EPA) accused

Volkswagen of devising a sophisticated scheme to deceive environmental authorities when testing for

nitrogen oxide (NOx) emissions. The notice of violation, and Volkswagen’s subsequent admission,

translated into an immediate 20% drop in the stock market value of VW shares due to concerns

about the company’s credibility as well as an estimated $18 billion in fines. A likely outcome of

this episode is the effective disappearance of diesel vehicles from the American market for a second

time in two decades due to failure to meet emission standards.

We use the VW scandal as motivation to explore the implications of fuel emissions policy

on the automobile industry. Rather than focus on the American market, we instead investigate the

European market where diesels represent roughly half of all new car sales and the emissions policy

chosen by regulators targets greenhouse gas emissions (CO, CO2) but is lax on emissions connected

to smog and acid rain (NOx). As diesels produce little CO and CO2 but a lot of NOx, this policy

appears to favor these vehicles. Our objective then is to assess whether European emissions policy

drove the diesel’s rise in Europe.

Using automobile registration data from Spain – a country which exhibited diesel adoption

rates representative of Europe as a whole – we estimate a discrete choice oligopoly model similar

to Berry, Levinsohn and Pakes (1995), henceforth BLP, to study an industry which is far from

competitive and where products are horizontally differentiated. The BLP framework has become

a workhorse model in the empirical Industrial Organization literature as it is flexible enough to

generate reasonable substitution patterns between similar products while accounting for product

characteristics known to consumers and firms but not to the researcher. For our purposes, the

estimated model provides a laboratory to explore the equilibrium effects of more rigorous NOx

emissions policies which we model as an increase in the marginal costs of production required to

reduce diesel NOx emissions. We call these costs “retrofitting costs.”

We find that diesels generated significant profits for European firms and the success of

diesel vehicles in Europe was, in large part, the consequence of diesel-friendly emissions standards

employed by European regulators. The model predicts that for even moderate levels of retrofitting

costs firms maximize profits by increasing diesel prices and consumers respond by shifting con-

sumption towards fuel-efficient gasoline engines produced by foreign auto makers. In other words,

had the EU chosen a more rigorous NOx standard, the popularity of diesels and the inherent

competitive advantage they provided domestic auto makers would have decreased significantly,

leading to an increase in imports from primarily Asian car manufacturers. Only by increasing the

import tariff from 10.3% to between 13−16% could EU regulators have pushed import penetration

– 1 –

back to the level observed under the current EU emissions policy. This indicates the pro-diesel EU

emissions policy amounted to a significant non-tariff trade policy as it had a disproportionately

positive impact on those firms which had invested to offer diesels, namely domestic automakers.

A unique feature of our data set is that it captures the rapid consumer adoption of next

generation diesel engines. Diesels were now significantly quieter, cleaner (i.e., no black smoke), and

more reliable than their predecessors while maintaining superior fuel efficiency and torque relative to

comparable gasoline models. Our structural model then enables us attribute the diffusion of diesel

vehicles to observable characteristics such as price and unobservable factors such as customers

learning about these next generation diesel vehicles. Our estimates indicate that customer learning

did play a role in the diffusion of diesels, therefore imposing more rigorous NOx standards early in

the diffusion of this innovation not only would have limited contemporaneous diesel sales but also

would have decreased future demand for diesels. Further, we find this channel has large quantitative

implications as it nearly doubles the equivalent import tariff implied by EU emissions policy (from

13% to 24%).

Researchers have historically focused on the role of explicit trade policies in shaping in-

dustries. An important example in the automobile industry is the voluntary export restraints

placed on Japanese cars during the 1980s and early 1990s. Feenstra (1988) documents significant

quality-upgrading by Japanese firms leading to the growth of luxury brands Acura, Infiniti, and

Lexus in the US market. Berry, Levinsohn and Pakes (1999) show this policy increased profits for

domestic firms and decreased welfare for domestic consumers while leaving significant tariff revenue

on the table.

While the most popular tool to distort trade flows has been historically import tariffs,

multilateral negotiations have largely driven these to zero. The result has been a growing empirical

literature documenting the effects of trade liberalization on firm behavior. For example, Edmond,

Midrigan and Xu (2015) evaluate the competitive effects of international trade on the Taiwanese

electronics industry finding that foreign competition decreases misallocation and markups. Aw,

Roberts and Xu (2011) find empirical evidence that international markets also provide incentive

for firms to innovate.

Our results are important as they provide evidence that in a world with ever more free trade

agreements, national policies such as environmental regulations can be effective tools to promote

local manufacturers over competitive imports, e.g., Ederington and Minier (2003) – perhaps our

most original contribution. We also show that a non-tariff domestic policy which promotes the

adoption of domestic innovations (or the development of innovations uniquely preferred by domestic

consumers) can have a significant effect on trade by influencing demand broadly while tariffs just

– 2 –

target price. This is a novel insight since such a policy may be welfare improving whereas tariffs

are generally thought to be welfare decreasing.

To the best of our knowledge, this paper is the first application of an equilibrium model

commonly used in empirical industrial organization to provide evidence of rent-seeking by countries

using domestic policy. As such, it builds on the literature related to the interaction of domestic

policy and international trade. The seminal theoretical contribution on this topic is Bhagwati and

Ramaswami (1963) who address the substitutability between domestic policy and import tariffs.

More recent works (e.g., Staiger 1995, Bagwell and Staiger 2001, Deardorff 1996, Thurk 2014) take

a more game theoretic approach and show that countries can use their domestic policies to extract

rents from the rest-of-the-world leading to a suboptimal aggregate outcome.

To be clear, we are not making a statement about Pareto optimality nor are we claiming that

European regulators designed their emission standards strategically to explicitly promote domestic

auto makers. Rather, we argue that regardless of whether it was the intent of the policymaker

or not, the effect of the environmental policy was to protect domestic European auto makers by

encouraging a “home bias” for domestic consumers. We show our conclusions are robust to a

variety of different assumptions and specifications but perhaps the most telling support are the

EU Commission’s own actions after Volkswagen announced it had also been cheating on European

emissions policies since 2004. Rather than penalizing the firm with substantial fines or removing

its cars from the European marketplace, EU regulators chose to increase the NOx ceiling facing

cars sold in Europe and committed to not revisiting the policy until 2019. We view this as evidence

that the NOx emissions policy employed in Europe was intimately related to the health of at least

Volkswagen, if not all European auto makers.

Finally, the paper contributes to the current discussion on identification and estimation

of BLP models by showing the practical repercussions of ignoring potential correlation between

observable and unobservable product attributes. An important identification assumption used

in a standard BLP estimation is that observable and unobservable product characteristics are

uncorrelated. Given the paucity of observable characteristics – a common trait of BLP models –

it seems plausible that unobservable characteristics are quantitatively important factors in deter-

mining consumer purchases as well as correlated with the observable characteristics either in the

cross-section or across time. This is particularly true for us since diesels comprise a significant share

of the market and many of the key features determining demand such as reliability, durability, and

high torque at low r.p.m. are unobservable to econometricians but likely related to the observed

weight or mileage as diesel vehicles are heavier and more fuel-efficient. Assuming exogeneity,

therefore, may introduce quantitatively significant biases into the estimation and subsequent policy

experiments – a hypothesis which we test.

– 3 –

Our estimation approach follows Petrin and Seo (2016) and allows for correlation between

observable and unobservable automobile characteristics using the firms’ first-order conditions for

profit-maximization as moment conditions – an approach similar to Hansen and Singleton (1982).

The idea is straightforward. Each period firms choose product attributes, observed or otherwise,

to maximize profits. Though firms may have different beliefs or information regarding competitors’

attribute choices, they all understand that their attribute choices influence equilibrium prices of

automobile manufacturers through own and cross-price effects. The first-order conditions implied by

the Bayes-Nash equilibrium provide moment conditions to estimate the structural demand and cost

parameters. Intuitively, a firm’s choice to make both larger (observable to us) and more reliable cars

(unobservable to us) provides information about both consumer preferences and production costs

for both attributes even though vehicle reliability, for example, can only be inferred by consumer

purchases conditional on observable characteristics.

Perhaps not surprisingly, we do find significant correlation between the product characteris-

tics we observe and the unobserved product characteristics needed to reconcile consumer purchases.

We also find that our estimation approach yields demand estimates which are more elastic than

employing a standard BLP estimation as well as more reasonable and significant point estimates

– a similar finding to Petrin and Seo (2016). Ignoring the correlation between observable and

unobservable product characteristics in the estimation would have led us to significantly over-state

estimated markups and the profitability of diesels to European firms. The effect on the implicit

import tariff is less stark, however, particularly for small retrofitting costs. We view this as both

reinforcing our main results and providing some consolation that the bias introduced when a

researcher estimates demand using the much simpler approach outlined in BLP may be small.

The paper is organized as follows. In section 2, we describe Volkswagen’s introduction

of the tdi innovation, its imitation by European auto makers, and summarize the main features

of the Spanish market for diesel automobiles. In Section 3 we document differences in emissions

policy between the US and EU; differences which materialized in the Volkswagen scandal. Section 4

describes the equilibrium model of discrete choice demand for horizontally differentiated products.

Section 5 describes the estimation approach, discusses identification, and reports the estimation

results in comparison to those implied by the standard BLP estimation strategy. In Section 6

use the estimated model to quantify the equilibrium implications of alternative emissions policy

on the European automobile industry. In Section 7 we evaluate the implications to our results of

assuming product exogeneity. Finally, Section 8 summarizes our results and contribution as well

as discusses avenues for future research. Details of the estimation, additional results, data sources,

and institutional details of the Spanish automobile market are documented in the Appendix.

– 4 –

2 The European Market for Diesel Automobiles in the 1990s

This section familiarizes the reader with the basic characteristics of the diesel technology; the

institutional features of the European market that allowed for a swift take off of diesel sales in the

early 1990s; and the evolution of the Spanish market.

2.1 A Significant Innovation - Next Generation Diesel Engines

In the late 19th century, Rudolf Diesel designed an internal combustion engine in which heavy fuel

self-ignites after being injected into a cylinder where air has been compressed to a much higher

degree than in gasoline engines. However, it was only in 1927, many years after Diesel’s death, that

the German company Bosch built the injection pump that made the development of the engine for

trucks and automobiles possible. The first diesel vehicles sold commercially followed soon after: the

1933 Citroen Rosalie and the 1936 Mercedes-Benz 260D. Large passenger and commercial diesel

vehicles were common in Europe from the late 1950s through the 1990s.

In 1989, Volkswagen introduced the tdi diesel engine in its Audi 100 model, a substantial

improvement over the existing Perkins technology. A tdi engine uses a fuel injector that sprays

fuel directly into the combustion chamber of each cylinder. The turbocharger increases the amount

of air going into the cylinders and an intercooler lowers the temperature of the air in the turbo,

thereby increasing the amount of fuel that can be injected and burned. Overall, tdi allows for

greater engine performance while providing more torque at low r.p.m. than alternative gasoline

engines. They are also credited with being more durable and reliable than gasoline engines

although this was something yet to be learned by consumers at the time this technology was first

introduced.1 Following this major technological breakthrough, European manufacturers other than

volkswagen improved their diesel engines and European drivers enthusiastically embraced diesel

automobiles. The incredible pace of adoption of diesel automobiles suggests that the tdi proved

to be a significant technological advance and consumers gained little from waiting for additional

incremental improvements, which have been few and of minor importance.2

2.2 Initial Market Conditions

There are important institutional circumstances that helped build the initial conditions that were

particularly favorable for the adoption of this new technology in Europe. The key element triggering

1 See the 2004 report “Why Diesel?” from the European Association of Automobile Manufacturers (ACEA).2 This argument was first put forward by Schumpeter (1950, p.98) and later formalized by Balcer and Lippman

(1984). More recently, it has been used by Manuelli and Seshadri (2014) to explain the half a century time spanneeded for the diffusion of the much studied case of tractors.

– 5 –

all these favorable development is the European Fuel Tax Directive of 1973. Following the first

oil crisis of 1973, the then nine members of the European Economic Community gathered in

Copenhagen in December of that year and agreed to develop a common energy policy. A main idea

was to harmonize fuel taxation across countries so that drivers, and fossil fuel users in general, faced

a single and consistent set of incentives to save energy. Coordination also limited the possibility of

arbitrage across state lines as well as some countries free riding on the conservation efforts of other

members. Fuel prices or their taxation were not harmonized overnight but the new Tax Directive

offered principles of taxation that were eventually followed in every country. For the purposes of

this study, the two most prominent features of this Directive are that fuels are taxed by volume

rather than by their energetic content and that diesel fuel is taxed at a lower rate than gasoline.

Figure 1 shows that in our sample diesel tax amounted to about 69% of gasoline tax (32 vs. 46

Euro cents per liter) resulting in systematically lower prices for diesel fuel.

Figure 1: Fuel Prices Gross and Net of Taxes (1994 Eurocents/liter)

0

10

20

30

40

50

60

70

80

1992 1993 1994 1995 1996 1997 1998 1999 2000

Fuel Price Plus Tax

(a) Gasoline

0

10

20

30

40

50

60

70

80

1992 1993 1994 1995 1996 1997 1998 1999 2000

Fuel Price Plus Tax

(b) Diesel

Taxing fuels by volume offers a transparent criteria to monitor national policies. However,

it also creates an incentive to use diesel fuel as diesel engines consume less per mile due to its higher

energy content (129,500 BTU per gallon vs. gasoline’s 114,000). The favorable tax treatment of

diesel fuels exacerbated this effect. Historically, this approach was intended partly to help two

economic industries particularly hit by the increase in oil prices: road transport and agriculture.

With minor modifications, these principles have guided European fuel taxation until very recently.

In 1997 the European Commission first suggested modifying these principles of taxation to reduce

the differential treatment of diesel and gasoline fuels and incorporating elements of environmental

impact of each type of fuel when setting taxes. It should be noted that this change in principles

– 6 –

were only adopted in 2013. Thus, consumers faced stable and consistent incentives favoring diesel

fuel consumption for a very long period of time.3

This favorable tax treatment of diesel fuel fostered the sale of diesel vehicles from the

mid-1970s on. By the end of the 1980s, some large passenger cars and many commercial vehicles

comprising almost 10% of the market ran on diesel fuel. Thus, when the tdi was first sold in 1989,

Europeans, unlike Americans, were familiar with diesels and did not have a particularly negative

perception of the quality of diesel vehicles.4 More importantly, Europeans did not have to cope with

the additional network costs commonly delaying the adoption of alternative fuels: by 1990 diesel

pumps were ubiquitous, indeed available in every gas station, and it was easy to find mechanics

trained to service these vehicles in case repairs were needed.

Initial conditions were thus more conducive to the success of the tdi technology than in

any other automobile market in the world. And yet, it was not obvious that consumers were

going to end up embracing this new technology when volkswagen introduced the tdi engine.

Diesels are known to achieve better mileage than otherwise identical gasoline vehicles, leading to

future fuel cost savings, but they are also more expensive to purchase, presumably due to higher

production costs or because manufacturers’ attempt to capture consumer rents of drivers favoring

diesel vehicles.5 But since the diffusion of tdi coincided with a long period of historically low and

stable fuel prices documented in Figure 1, the value of potential fuel savings were limited and so

was the manufacturers’ ability to overprice diesel automobiles.

Figure 2 demonstrates that imitation of the tdi occurred quickly and was largely driven

by rival European auto makers. This indicates the ineffectiveness of volkswagen to protect its

innovation via the patent system since it was not difficult for rival firms to offer their own improved

diesel models in a relatively short period of time. This suggests a low cost of imitation – a trait which

characterizes all “general technologies” (e.g., Bresnahan 2010) – due to the fact the technology can

be easily modified or reverse engineered. For consumers, imitation led to the introduction of more

variety and better quality of vehicles to choose from while competition intensified in this market

3 See http://ec.europa.eu/taxation_customs/taxation/excise_duties/energy_products/legislation/index_

en.htm for a complete description of the European Fuel Tax Directive and its evolution over time.4 See http://www.autosavant.com/2009/08/11/the-cars-that-killed-gm-the-oldsmobile-diesel/ for an ac-

count of how badly gm’s retrofitted gasoline engines delivered poor performance when running on diesel fuelin the late 1970s and early 1980s and how such experience conformed the negative views of Americans on dieselvehicles for many years.

5 Verboven (2002) studies the price premium paid for diesel vehicles relative to otherwise identical gasoline modeland explains it as business strategy aimed to capture some of the rents of consumers with heterogeneous drivinghabits.

– 7 –

Figure 2: Number of Diesel Models Offered

0

10

20

30

40

50

60

70

80

90

100

1992 1994 1996 1998 2000

VW PSA Other EU Foreign

segment, keeping diesel vehicles affordable, and therefore reinforcing the fast diffusion of this new

technology.6

2.3 Evolution of Automobile Characteristics

Our data include yearly car registrations by manufacturer, model, and fuel engine type in Spain

between 1992 and 2000. After removing a few observations, mostly of luxury vehicles with extremely

small market shares, our sample is an unbalanced panel comprising 99.2% of all car registrations in

Spain during the 1990s.7 Spain was the fifth largest automobile manufacturer in the world during

the 1990s and also the fifth largest European automobile market by sales after Germany, France,

the United Kingdom, and Italy. In our sample automobile sales range from 968,334 to 1,364,687

units sold annually.

Figure 3 documents the evolution and composition of sales in Spain during the 1990s.

Figure 3(a) shows that sales of gasoline models were essentially identical in 1993 and 1995, about

573,000, despite a scrappage program in 1994, when they temporarily increased by 15%. Since

then, sales of gasoline models increased at a steady pace until 1999 but they never reached the 1992

peak level again. The evolution of sales of diesel automobiles is starkly different. Initially in 1992,

they only represented 16% of total sales but by the end of the decade diesels represented 54% of the

6 volkswagen was an important firm but not the leader in the Spanish market: renault was by far the leader inthe gasoline segment and psa, which includes the French brands citroen and peugeot, in diesel. See Figure A.1in the Web Appendix.

7 See Appendix A for further details.

– 8 –

Figure 3: Automobile Sales and Models by Year and Fuel Type

0

75000

150000

225000

300000

375000

450000

525000

600000

675000

750000

825000

91 92 93 94 95 96 97 98 99 00

Automobile Sales

Gasoline Diesel

(a) Sales (thousands of vehicles)

0

20

40

60

80

100

120

140

91 92 93 94 95 96 97 98 99 00

Number of Models

Gasoline Diesel

(b) Models Available

market, growing from 161,667 to 732,334 units sold in years 1992 and 2000, respectively. Moreover,

diesel adoption in Spain mirrored that of other European countries.8

There was an equally impressive transformation of supply to meet this quick shift in

demand. Figure 3(b) shows that by 1992, manufacturers already offered 44 diesels out of 141

models sold (although not all of them comparable to tdi). Such a large number of diesel models

hints at automobile manufacturers fearing business stealing much more than the consequences of

cannibalizing the sales of their own gasoline models. It also suggests that volkswagen expected

competitors to enter this segment when it decided to introduce the tdi technology. Furthermore,

the number of models available grew significantly, both in the gasoline and the diesel segments,

reflecting the effective entry of Asian manufacturers in the European market and a substantial

increase in competition among fuel-efficient vehicles.9 Since the entry of new models should reduce

markups, consumers benefited from both an increase in variety and lower prices.

Table 1 summarizes the evolution of the features of vehicles sold in the Spanish automobile

market during the 1990s.10 By the end of the decade auto makers produced cars which were 36.1%

more expensive, are 4.4% larger, and 11.6% less powerful (i.e., hpw). The combined effect are

cars that are 2.3% less fuel-efficient in terms of mileage and 2.8% less expensive to drive when we

account for increasing fuel prices.

We document even more dramatic changes across gasoline and diesel models. While the

prices of gasoline and diesel models both increased over the decade, the increase for gasoline

8 See Figure A.2 in Appendix A.9 Asian imports include daewoo, honda, hyundai, kia, mazda, mitsubishi, nissan, suzuki, and toyota.chrysler is the only non-Asian imported brand. Thus, we use the terms “Asians” or “non-Europeans” whenreferring to imports. chrysler sold its production facilities to peugeot in 1978 and since then the few modelssold in Europe are imported from the United States. On the contrary ford and gm are considered Europeanmanufacturers. ford has 12 manufacturing plants and has been continuously present in Europe since 1931. gmentered the European market in 1911, acquired the British brand Vauxhall and the German Opel in the 1920s andtoday operate 14 manufacturing facilities in Europe.

10Table A.2 in Appendix F.1 complements this description of product features reporting statistics by market segment.

– 9 –

Table 1: Car Model Characteristics by Origin and Engine Types

year/group models share price c90 kpe size hpw

1992

eu: diesel 43 16.60 12.26 4.45 46.42 73.84 3.14eu: gasoline 73 79.45 11.05 5.39 29.62 71.50 4.12non-eu: diesel 1 0.09 13.76 5.30 38.58 80.51 2.86non-eu: gasoline 24 3.86 14.88 5.82 27.31 77.99 4.53

all 141 100.00 11.40 5.25 32.33 72.15 3.97

2000

eu: diesel 75 50.95 16.19 4.55 38.18 76.32 3.14eu: gasoline 84 37.28 14.93 5.68 24.23 73.40 3.90non-eu: diesel 20 2.71 17.20 5.41 32.63 82.48 3.22non-eu: gasoline 50 9.06 13.66 6.11 22.80 75.32 4.08

all 229 100.00 15.52 5.13 31.43 75.31 3.51

Statistics weighted by relevant quantity sold. share is the market share as defined by automobiles sold.price is denominated in the equivalent of thousands of 1994 Euros and includes value added taxes andimport tariffs. c90 is consumption (in liters) of fuel required to cover 100km at a constant speed of 90km/hr. kpe is the distance, measured in kilometers, traveled per euro of fuel. size is length×widthmeasured in square feet. hpw is the performance ratio of horsepower per hundred pounds of weight.

models was much larger (46.5% vs 27.7%). Interestingly, this coincided with the transformation

of European production in just a few years: European vehicles represented 96% of sales at the

beginning and 88% at the end of the 1990s. But while only less than one out of five European cars

was a diesel in 1992, by year 2000 they sold four diesels for every three gasoline models. Overall, a

quarter of a million fewer gasoline vehicles were sold by the end of decade while the production of

diesel models increased by over half a million units, almost quadrupling production. Sales of diesel

became so important in the market that non-European auto makers began introducing their own

diesel models.

When deciding what type of engine to purchase, drivers compare observable product char-

acteristics as well as likely related expected performance attributes of each engine, unobservable

to econometricians. Since the difference between a diesel and gasoline version of a particular car

model depends on only what’s under the hood (i.e., they share the same chassis), a consumer

deciding between aa Audi A4 gas or diesel car bases her decision on differences in performance not

car size. Specifically, diesel vehicles are about 10% heavier than similar gasoline versions; have 15%

to 20% less horsepower than gasoline vehicles; and are between one and two thousand euros more

expensive. Finally, diesel vehicles consume 20%− 40% less fuel than a comparable gasoline model,

enabling a diesels to travel about 63% farther on a euro of fuel.

– 10 –

Figure 4: Change in the Distribution of Automobile Attributes

(a) Gasoline: Mileage (c90) (b) Diesel: Mileage(c90)

(c) Gasoline: Cost of Driving (kpe) (d) Diesel: Cost of Driving (kpe)

(e) Gasoline: size (f) Diesel: size

(g) Gasoline: Performance (hpw) (h) Diesel: Performance (hpw)

– 11 –

For diesels to succeed as they did, it is likely that this new technology was seen as desirable in

many ways, and not only regarding fuel economy. The shift in the distributions of some observable

automobile characteristics is shown in Figure 4 and formal tests of first and second order stochastic

dominance are presented in Table F.1 in Appendix F.1. Despite the fact that all vehicles became

larger, heavier and slightly more powerful during the decade, there is little evidence that gasoline

vehicles differ much during the 1990s. Kolmogorov-Smirnov tests indicate that neither the early or

late distribution of attributes of gasoline models dominate each other with the exception of kpe;

the cost of driving gasoline vehicles is definitely higher by year 2000 (as a consequence of increasing

fuel prices). Diesel vehicles on the other hand, show sign of substantial change during the decade;

they are also more expensive to drive (kpe) by year 2000 despite the fact that they became more

fuel-efficient (c90), as they are also larger (size) and show weakly better performance (second

order stochastic dominance in hpw). All this descriptive evidence hints at diesel vehicles becoming

better products capable of increasingly attracting the interest of many drivers.

3 Vehicle Emissions Standards in the United States and Europe

Thus far we have documented the popularity of diesels among consumers (Figure 3) and discussed

the significant imitation of diesels by largely European firms within a few years of Volkswagen’s

introduction of the tdi (Figure 2). And yet, despite how easy imitation appears to be, diesels almost

disappeared in the U.S. during the same period of time. The common explanation for the different

evolution of these two large markets attributes the success of diesels in Europe to the favorable

tax treatment of the diesel fuels in Europe. This is a popular explanation that lacks empirical

support, however. While it is true that reduced taxation of diesel fuel favors larger penetration of

diesel vehicles in a cross-section of mature markets, e.g., Grigolon, Reynaert and Verboven (2015),

fifteen years of such policy only led to a 10% market share penetration by the early 1990s. This

suggests that preferential diesel fuel taxes played a minor role in promoting diesel adoption prior

to the tdi.11

In this section we put forward the novel hypothesis that the different fate of diesels in

Europe and the U.S. was instead due to the different goals pursued by the environmental policies

in the U.S. and in Europe (Figure 5).12 While Americans were concerned mostly with reduction in

emissions leading to acid rain, Europeans aimed at reducing green house emissions.

11 In Appendix E we use the estimated model to show the effect after the tdi’s introduction was also small.12European authorities set NOx and particulate matter (PM) standards for each vehicle while U.S. authorities set

a fleet-wide limit. As for CO and CO2 emissions, these depend on fleet average fuel consumption standards andare reported in Figure 5 as realized fleet-wide levels. See Section IV of the 2001 report: “Demand for Diesels: TheEuropean Experience. Harnessing Diesel Innovation for Passenger Vehicle Fuel Efficiency and Emissions Objectives”available at www.dieselforum.org.

– 12 –

In the United States, the approval of the 1990 Clean Air Act Amendments (CAAA) directed

the U.S. Environmental Protection Agency (EPA) to, among many other things, reduce acid rain

produced by nitrogen oxide (NOx) and sulfur dioxide (SO2). The EPA therefore chose a policy

largely aimed at power generating plants which set emission reduction goals (Title IV-A) and

established a cap-and-trade system (Title V), but it also translated into an ever more stringent

NOx emission standards for light-duty vehicles (Title II-A).

Figure 5: Europe and U.S. Emissions Standards

0.07

0.4

0

0.1

0.2

0.3

0.4

0.5

0.6

0.7

0.8

0.9

1

US EU

NOx

0.01

0.04

0

0.02

0.04

0.06

0.08

0.1

0.12

0.14

0.16

0.18

0.2

US EU

PM

2.89

0.8

0

0.5

1

1.5

2

2.5

3

3.5

4

US EU

CO

330

224

0

50

100

150

200

250

300

350

400

US EU

CO2

Source: www.dieselforum.org. All statistics are for the year 2000 and are in grams per mile. “NOx” refers tonitrogen oxide limits; “PM” to particulate matter; “CO” carbon monoxide; and “CO2” carbon dioxide.

European regulators took a different approach and chose a less stringent emission standard

on NOx and PM (Figure 5).13 While in 1994 U.S. Tier 1 standard allowed NOx emissions of 1

gram per mile (g/mi), the Euro I standard was 1.55g/mi. By year 2000, the U.S. policy allowed only

0.07g/mi while the Euro III standard set the NOx emission level at a far less demanding 0.4g/mi

level. Similar results hold for PM . The fast diffusion of diesel vehicles in the 1990s likely also

enabled European authorities to choose more stringent CO2 emission standards than the United

States; the goals of local automobile manufacturers and European environmental regulators were

thus perfectly aligned. Were these differences in environmental goals enough to explain the different

evolution of diesels in the U.S. and Europe? Absent any data on sales of automobiles by type of

13The negative health effects of PM are well documented. Capturing PM is however easier and far lessexpensive than capturing NOx and we will not address it in our counterfactual analysis. See The WorldBank’s report: Reducing Black Carbon Emissions from Diesel Vehicles: Impacts, Control Strategies, andCost-Benefit Analysis available at https://openknowledge.worldbank.org/bitstream/handle/10986/17785/

864850WP00PUBL0l0report002April2014.pdf. In page 27 it indicates that the cost of complying with the moststringent PM emissions for a 4-cylinder 1.5 L diesel engine was $1, 400 in 2014.

– 13 –

engine in the American market, we argue that this is the case based on anecdotal evidence for the

U.S. and our counterfactual analysis for Europe.

The differences between the U.S. and European standards are significant for automobiles

since reducing NOx emissions is much harder for diesel engines as the three-way catalytic converters

used to reduce emissions in gasoline engines cannot cope with the high concentrations of NOx

generated by diesel engines (e.g., Canis 2012). Thus, rather than investing to redesign their diesel

engines to meet these stringent emission standards, volkswagen and mercedes chose to stop

selling their diesel models in the U.S. market in 1993 and 1994, respectively, precisely at the time

of the implementation of the U.S. emission standards mandated by the CAAA.14 Only in 2010 did

the EPA finally address the issue of NOx emissions from diesel vehicles by requiring the installation

of an urea-based selective catalytic reduction that injects an aqueous solution into the vehicles’

exhaust stream to “scrub” NOx emissions. Since then, auto makers have introduced more diesel

models into U.S. market, including those states that adhere to the even more demanding California

emission standards. All these circumstances suggest that the imposition of these emission standards

amounted to a de facto ban of diesel vehicles in the U.S. market. Could then a similar European

emission policy have eliminated any chance of success for diesels in Europe?

4 An Equilibrium Oligopoly Model of the Automobile Industry

In this section we present a structural equilibrium model of demand and supply which we use to

discipline our analysis. We first present a standard BLP model of discrete choice demand with

heterogenous customers. Then we describe an oligopoly model of supply in which multi-product

firms comprising the several brands detailed in Table A.1 choose prices conditional on product

characteristics (e.g., car size).

4.1 Demand

Demand can be summarized as follows: consumer i derives an indirect utility from buying vehicle

j at time t that depends on price and characteristics of the car:

uijt = xjtβ∗i − α∗i pjt + ξjt + εijt ,

where i = 1, . . . , It; j = 1, . . . , Jt; t = 1992, ..., 2000 .(1)

14According to Stewart (2010), the NOx emissions level of the least polluting diesel model available in Canada,the volkswagen Jetta (known as Bora in Europe), was 0.915 and 0.927g/mi for the 1991 and 1997 year models,respectively. This indicates that the NOx emissions standards imposed by the EPA were indeed binding constraintsfor diesel vehicles since even the cleanest diesel models barely met the 1994 U.S. emission standards and wouldhave generated NOx emissions thirteen times greater than the 2000 limit.

– 14 –

This Lancasterian approach makes the payoff of a consumer depend on the set of characteristics of

the vehicle purchased, which includes a vector of K observable vehicle characteristics xjt as well as

others that remain unobservable for the econometrician, ξjt, plus the effect of unobserved tastes of

consumer i for vehicle j, εijt, which is assumed i.i.d. multivariate type I extreme value distributed.

We allow for individual heterogeneity in response to vehicle prices and characteristics by modeling

the distribution of consumer preferences over characteristics and prices as multivariate normal with

a mean that shifts with consumer attributes:15

(α∗i

β∗i

)=

(α

βt

)+ ΠtDit + Σtρit , ρit ∼ F . (2)

Consumer i in period t is characterized by a d vector of observed demographic attributes,

Dit, as well as a vector of random tastes, ρit distributed i.i.d. with cumulative distribution function

F which is commonly assumed to be standard normal. Πt is a (n+1)×d matrix of coefficients that

measures the effect of income on the consumer valuation of automobile characteristics, e.g., average

valuation and price responsiveness. Similarly, Σt measures the covariance in unobserved preferences

across characteristics. We decompose the deterministic portion of the consumer’s indirect utility

into a common part shared across consumers, δjt, and an idiosyncratic component, µijt. The mean

utilities of choosing product j and the idiosyncratic deviations around them are given by:

δjt = xjtβ + αpjt + ξjt , (3a)

µijt =(xjt pjt

)×(

ΠtDit + Σtρit

). (3b)

Consumers choose to purchase either one of the Jt vehicles available or j = 0, the outside

option of not buying a new car with zero mean utility, µi0t = 0. We therefore define the set of

individual-specific characteristics leading to the optimal choice of car j as:

Ajt (x·t, p·t, ξ·t; θ) = (Dit, ρit, εijt) |uijt ≥ uikt ∀k = 0, 1, . . . , Jt , (4)

with θ summarizing all model parameters. The extreme value distribution of random shocks allows

us to integrate over the distribution of εit to obtain the probability of observing Ajt analytically.

The probability that consumer i purchases automobile model j in period t is:

sijt =exp (δjt + µijt)

1 +∑k∈Jt

exp(δkt + µit). (5)

15Random coefficients generates correlations in utilities for the various automobile alternatives that relax the restric-tive substitution patterns generated by the Independence of Irrelevant Alternatives property of the logit model.

– 15 –

Integrating over the distributions of observable and unobservable consumer attributes Dit and ρit,

denoted by PD(Dt) and Pρ(ρt), respectively, leads to the model prediction of the market share for

product j at time t:

sjt(xt, pt, ξt; θ) =

∫ρt

∫Dt

sijtdPDt(Dt)dPρt(ρt) , (6)

with s0t denoting the market share of the outside option.

4.2 Supply

Equilibrium prices are found as the solution to a non-cooperative Bertrand-Nash game among the

competing auto makers. Specifically, equilibrium prices (pτj ) can be written a nonlinear function of

the product characteristics (x), market shares sj(x, p, ξ; θ), retail prices (p), and markups:

pτj = mcj + ∆−1(p, x, ξ; θ)sj(p, x, ξ; θ)︸ ︷︷ ︸bj(p, x, ξ; θ)

, (7)

where pj = pτj × (1 + τj) and τj is the import duty applicable to model j, if any. The vector of

equilibrium markups bj(·) depends on market shares sj(·) and the matrix ∆(·) with elements:

∆rj(x, p, ξ; θ) =

∂sr(x, p, ξt; θ)

∂pj× ∂pj∂pτj

, if products r, j ∈ Ff ,

0 otherwise .

(8)

In estimating costs we make a common assumption that firms have Cobb-Douglas cost

functions of the following (log-linear) form:

log cj =∑k

γk log(Xkj ) + γξξj + ηj︸ ︷︷ ︸

ωj

. (9)

Marginal costs are therefore a function of both observed and unobserved product characteristics,

via X and ξ, and an unknown (to the econometrician) cost component η. Explicitly modeling ξ

in the cost function does two things. First, it illustrates the potential endogeneity and subsequent

estimation bias in the supply-side estimation since movements in ξ will be captured in ω in any

standard BLP model. Second, it provides the structure to account for changes in unobserved

product attributes ξ on marginal cost, i.e., ∂c/∂ξ. This is particularly relevant in our case as many

features likely driving the cost of diesel vehicles such as torque, reliability, and durability remain

unobservable to us.

– 16 –

5 Estimation

The common estimation approach employed by researchers using the BLP framework is to assume

exogeneity of product characteristics so that E[ξ|X] = 0 and E[ω|X] = 0. One can then estimate

the demand and supply parameters (Σ,Π, β, γ) using observable product characteristics as basis

functions to construct identifying moment conditions.

In this section we describe an alternative estimation approach which follows Petrin and Seo

(2016). The advantage of this approach is that it allows us to estimate the key demand and supply

parameters with limited restrictions on the relationship between the observable (X) and unobserv-

able (ξ) product characteristics. To do so we extend the model described in Section 4 as follows:

each period multi-product automobile manufacturers choose the characteristics of their products

(both the K observed and the unobserved ξ), understanding that these characteristics determine

the retail prices via equation (7). In a Bayes-Nash equilibrium the chosen product characteristics

(and the ensuing retail prices) maximize profits conditional on each firm’s belief about the actions

of its rivals. Mathematically, each multi-product firm f chooses product characteristics xkj in period

t to solve:

maxxkjt

E

∑r∈Fft

(prt − crt)× srt(x, p, ξt; θ)∣∣∣Ψf

t

, (10)

where Fft is the set of vehicles of all brands sold by firm f and Ψft is its period t information

set which may include any information available to the firm in period t (e.g., observable and

unobservable product characteristics).16 To simplify notation going forward, we drop t subscripts

and will note when timing considerations are important. The subsequent optimal pricing strategy

will be a function of product positioning of all competing firms. Thus, in choosing product

attributes, profit maximization yields the following first order condition:

E

∂(pj − cj)∂xkj

× sj +∑r∈Ff

(pr − cr)×∂sr

∂xkj

∣∣∣Ψft

= 0 , (11)

16While allowing a firm to choose observable product characteristics may be intuitive, it is worthwhile to providesome intuition as to what it means for a firm to choose an unobservable product characteristic. For example, inequation (10) we allow a company such as audi to both increase the fuel efficiency (observable to us) and reliability(not observable to us) of a car in its portfolio. The latter choice is captured in the model via an increase in ξ, thoughof course audi could choose other factors which ultimately increase demand through ξ. Ignoring this relationshipleads to a violation of product characteristic exogeneity and biases estimation results. Of course, a cleverly chosenset of control variables would also remove any exogeneity concerns (e.g., Nevo 2000), but this is often difficult inpractice. An advantage of our approach is that we can be agnostic about the controls in the estimation and testfor exogeneity afterwards.

– 17 –

where:

∂sr

∂xkj=

∫ρk

∫D

(βk + σkρk + πkD)× sij(1− sir)dPD(D)dPρ(ρ) +∑m∈Ff

∂sr∂pm

∂pm

∂xkj, r = j,

−∫ρk

∫D

(βk + σkρk + πkD)× sijsirdPD(D)dPρ(ρ) +∑m∈Ff

∂sr∂pm

∂pm

∂xkj, otherwise .

(12)

In the BLP framework product attributes are taken as given even though they determine

pricing strategies and the ability to charge a higher or lower markups depending on the product

positioning of all firms. Profit maximization conditions (11)-(12) describe an alternative framework

where firms first choose product characteristics while taking into account the expected impact of

these choices on profits through retail prices facing consumers and the induced cross-price effects

on the demand of other products offered by the firm. Product attributes and prices are chosen

sequentially so firms do not respond by changing attributes to respond to prices as in a model

where prices and attributes were chosen simultaneously. Thus, product characteristics, observed or

unobserved, condition the optimal pricing strategies that are set in equilibrium.

5.1 Estimation Details

Our estimation must account for several important changes taking place during the 1990s such

as increasing personal income, reduction of import duties, and multiple mergers of automobile

manufacturers as well as differences in the information available to firms when they choose product

characteristics. When estimating the model we simulate individuals from yearly census data to

account for growth in income and the expansion of the Spanish economy (time-varying outside

option). Similarly, the marginal cost equation (Equation 7) controls for relevant import taxes

faced by manufacturers depending on their national origin and varies over the period considered.

To account for changes in firms’ product portfolios, we update matrix ∆rj every year to match

the ever changing ownership structure of this industry during the 1990s and correctly define the

multi-product first-order profit maximization conditions of the equilibrium model to be estimated.17

The final task is to map the firm information sets, Ψft , into pay-off relevant beliefs about

rivals actions. Unfortunately, there is no clear way to construct the mapping ex-ante as it is unclear

the degree to which firms in this industry are knowledgeable of the likely innovation decisions of

their rivals or of the tastes of consumers today conditional on their previous purchase decisions.

We instead take a simpler approach and impose these beliefs directly using two likely candidates.

Under “Model 1” we make a similar assumption to Hansen and Singleton (1982) and allow period t

17See Appendix A for details on acquisitions and mergers in the European automobile industry during the 1990s.

– 18 –

firms to know all period t− 1 product characteristics of their rivals. We also assume that firms can

observe (or at least infer) the unobserved characteristics (ξ) driving consumer purchase decisions.

The structural errors (ν) then are due to either these information (or belief) asymmetries or

measurement error. We treat this estimation as a likely conservative approach since the underlying

assumed information/ belief structure may reflect lags between product design and pricing once

vehicles are developed and sold. In “Model 2” we allow firms to know both the period t observable

and unobservable characteristics of their rivals. The interpretation behind the structural errors (ν)

then is that of measurement error.18 We view this estimation as a bound to identify the robustness

of our estimates to alternative belief structures where firms have near perfect foresight, i.e., the case

where details of product developments are common knowledge among automobile manufacturers.

5.1.1 The GMM Estimator As in the standard BLP estimation approach we estimate the struc-

tural parameters of the model by gmm as in Hansen and Singleton (1982). Define the parameter

vector θ = [β,Σ,Π, γ]. First, we solve for the mean utilities δ(θ) using the standard contraction

mapping outlined in Appendix I of BLP. Next we solve for the implied markups bjt and use price

data to construct marginal costs, assuming a pure strategy Bertrand-Nash equilibrium. In principle,

these first-order conditions of equation (11) are necessary for any Bayes-Nash equilibrium with

multiple interacting players, although meeting these conditions does not rule out the possibility of

multiple equilibria nor does it imply a restriction regarding the equilibrium selection mechanism

should multiple equilibria exist. We follow Bajari, Benkard and Levin (2007) and use the data –

the observed product, price, and consumption choices – as the equilibrium selection mechanism

which is sufficient provided the data is generated by a single equilibrium.

We construct the structural error νkj (θ) as follows:

νkj (θ) =∂[pτj − cj(θ)]

∂xkj× sj(θ) +

∑r∈Ff

[pτr − cr(θ)]×∂sr(θ)

∂xkj. (13)

The evaluation of the response of demand for all products of each firm to each change in product

characteristics makes this task particularly computationally-intensive as it requires solving repeat-

edly for the equilibrium price responses due to changes in product characteristics in addition to

evaluating numerically the multiple integrals of equation (12).19

Profit-maximization requires that in each period t the expectation of the structural error ν

conditional on the firm beliefs, Ψft , is zero for all products and characteristics, i.e., E[νkj,t(θ)|Ψ

ft ] =

18For example, using manufacturer’s suggested retail price (MSRP) rather than transacted price – a common featureof automobile data (and present in BLP) – likely introduces error into the pricing game which could materialize indeviations in the first-order conditions.

19We provide specific details of the solution algorithm in Appendix B.

– 19 –

0). Moreover, any function of variables in the information set are valid instruments so the set of

potential instruments is large. Following Newey (1990) we assume that:

Ω(θ) = E[ν(θ)′ν(θ)] , (14)

is a constant square matrix which defines the covariance structure of optimization errors where

ν(θ) = [νk=1·,t (θ), ..., νk=K+1

·,t (θ)] is a matrix of structural errors and K + 1 is the number of

endogenous characteristics chosen by firms.20 Chamberlain (1987) shows that we can use the

model to generate instruments defined below:

Hjt(θ) = E

[∂νjt(θ)

∂θ

∣∣∣Ψft

]′Ω−1 . (15)

where Hjt is N -by-(K+ 1) matrix with N corresponding to the number of elements in θ. The logic

behind these instruments is straightforward: they place relatively more weight on observations

that are responsive to deviations of the parameter vector in a neighborhood of the estimated value.

Since the value of the structural error ν is dependent upon the assumed information structure, so

are the instruments.

We estimate θ using a commonly employed two-step gmm estimation. Specifically, in each

step we solve for the value of the GMM objective function conditional on θ by interacting the

structural errors (13) with the identifying moment conditions (15) as follows:

θ? = argminθ

G(θ)A−1G(θ)′ , (16)

where G(θ) ≡ E[ν(θ)′ × H(θ)], ν(θ) = [νk=1(θ); ...; νk=K+1(θ)] is a stacked vector of structural

errors, and A−1 is a positive-semidefinite weighting matrix that exists because there are (K + 1)

instruments for each element of θ, i.e.,

H(θ) =

H1(θ) . . . 0

... Hk(θ)...

0 . . . HK+1(θ)

. (17)

The estimator exploits the fact that at the true value of parameters θ?, the instruments

H are orthogonal to the errors ν(θ?), e.g., E[H ′ × ν(θ?)

]= 0. In the initial step we solve for H

and A−1 using the parameter estimates when we assume product characteristic exogeneity (i.e., a

standard BLP estimation). In constructing the weighting matrix, we allow for the structural errors

20See Newey (1993) for the case of heteroscedasticity.

– 20 –

ν within a car model to be correlated across characteristics and time. We then update θ by solving

(16) and using this value to update the instrument and weight matrices. We found additional

updates changed our estimates little.21 To ensure robustness of our gmm results we employed a

state-of-the-art estimation algorithm (KNITRO) shown to be effective with this class of models;

considered a large variety of initial conditions; and used the strict inner-loop convergence criterion

for calculating the mean utility δ suggested by Dube, Fox and Su (2012).

As a final step we identify systematic trends in estimated unobserved demand, ξ, and cost,

η, by projecting these vectors onto a set of dummy variables (e.g., diesel, segment, firm) and time

trends. We view these results as descriptive as many factors could be driving why consumers prefer

European cars to foreign cars or why alfa romeo is, ceteris paribus, more expensive to produce

than renault.

5.1.2 Specification Consumer demand (both mean and idiosyncratic) includes measures of auto-

mobile performance: horsepower divided by weight (hpw) and exterior dimensions (size). We also

include the fuel cost of driving, (kpe), as a “random coefficient” but we assume the distributions

for ρkpe is distributed i.i.d. exponential, therefore ρkpe plays a dual role, controlling not only the

mean valuation but also the substitution pattern within fuel (in)efficient vehicles. The advantage

of this approach is that ex-ante all agents value fuel-efficient cars (i.e., will value paying less to

get from point A to point B) though we allow agents to be heterogenous in how much they value

fuel-efficient cars (e.g., some agents may value producing less emissions of some kind, e.g., CO2 or

NOx). We also include random coefficients for diesel to generate different substitution within the

diesels and a constant to capture changes in substitution patterns due to the increasing product

set. We limit the demographic interactions (Π) to be just an interaction between price and income

as described in Section 4.1.

On the supply side, we include the logged values of the observed product characteristics

(hpw, size, and kpe) with some modifications. We replace kpe, which includes the effect of

fluctuations in the price of oil, with a measure solely based on fuel-efficiency, c90. Consequently,

audi’s choice of fuel-efficiency for a gasoline model A4 impacts its cost directly as measured by c90,

but demand for A4’s will also be influenced by changes in the price of gasoline due to economic

factors outside of audi’s control. Hence, we include kpe in the demand rather than in the supply

equation. Similarly, we allow for increasing steel prices to impact the cost of producing larger,

heavier cars by multiplying car weight and size by an index for the price of steel. This leads to

shifts of hpw and size in supply but not demand.

21 In principle one could make the problem just-identified by stacking the Hk(θ) matrices. We found this approachled to unrealistic variation in marginal costs, particularly across brands.

– 21 –

As mentioned above, the gmm estimation generates estimates of unobserved demand, ξ,

and supply, ω, which we project onto a set of dummies to identify systemic patterns in demand

and cost. As our sample period covers the diffusion of diesel vehicles, we include a diesel dummy

as well as a linear time interaction with diesel in order to capture the evolution of preferences

in favor of the new technology. We also include segment dummies allow for differences in demand

between, for example, minivans and luxury cars. Finally, we include a non-eu dummy to account

for differences in valuation of largely Asian imports in the European market and a dummy variable

for the national brand, seat, to evaluate possible home bias effects.

5.1.3 Parameter Identification This model represents a complex, nonlinear mapping from param-

eters to data, though the intuition behind our estimation approach is straightforward. In the model

consumers maximize utility by evaluating a car based not only price but also product characteristics,

some of which are unobserved to the researcher but known to consumers and firms. In Section 2.3

we documented that there is variation in product characteristics over the decade. Our estimation

then uses the firms’ choices in product characteristic space to reveal the underlying demand and

cost parameters via the correlation between the unobserved shocks (in both demand and cost) and

observable product characteristics. For example, if car size is positively correlated with reliability

and the latter increases consumer utility, volkswagen’s choice to produce large, reliable cars

provides information regarding the utility associated with car size in our estimation whereas the

two are assumed exogenous in the standard BLP approach leading to a biased estimator.

While there is no clear one-to-one mapping between a parameter and a specific moment in

the data, the intuition into how data variation identifies different components of θ is as follows.

Variation in the product set, product characteristics (e.g., size), prices, and quantities identifies

the random coefficients, Σ. We also include a random coefficient for diesel which is identified

by variation in the other moment conditions by fuel type. A similar argument holds for the

constant random coefficient which is identified by variation across moment conditions as the number

of products increase.22 Finally, the Bertrand-Nash pricing equilibrium plus variation in price

elasticities conditional on product characteristics identifies marginal costs, γ.

The intuition behind the identification of α (i.e., Π) is embedded in the structural error

ν(θ) – Equation (13) – where the key tension is between the objects (p− c) and∂[pτj−cj(θ)]

∂xkj× sj(θ).

22Note that a firm’s choice to offer a model with a diesel engine is a discrete choice while the structural errors impliedby the first-order conditions, equation (13), are based on continuous characteristics. An alternative approachwould be to redefine ν to include moment inequalities for diesel engines as in Pakes, Porter, Ho and Ishii (2015).Unfortunately, identifying the cost of introducing a diesel engine is difficult since automakers usually introduce adiesel engine option soon after launching a new gasoline car model (rather than introducing a diesel engine optionin an existing car model) so constructing these moment inequalities may not be informative. Since our model isover-identified we chose a simpler approach and ignored this endogenous choice (i.e., do not include a momentinequality on the optimality of offering car j as diesel).

– 22 –

As α ↑ 0 consumers become less price-sensitive and the markups implied by our Bertrand-Nash

pricing equilibrium increase, driving down estimated marginal costs and increasing (p − c). This

effect is also true in BLP. The innovation here then is the second term,∂[pτj−cj(θ)]

∂xkj× sj(θ), which

becomes large as consumer demand becomes more inelastic. As this term reflects the sensitivity of

price to changes in product characteristics, allowing for product characteristic endogeneity enables

the researcher to use changes in product characteristics to also infer market-power via α. As we

will see below, adding this additional level of identification leads to demand estimates which are

much more elastic.

5.2 Estimation Results

Estimation results are presented in Table 2. Overall, the estimates are reasonable, statistically

significant, and congruent with the descriptive evidence of the industry of Section 2. We also

observe little difference between Models 1 and 2 suggesting the information sets we considered

played only a minor role. For brevity, we will largely refer to the estimation results from Model 1.

Starting with costs, diesels are more expensive to manufacture than gasoline models. Marginal

cost of production are also higher for larger and more powerful cars. Marginal cost is decreasing

in fuel efficiency (increasing in c90) in both specifications of our full model, though the effect

is small and not statistically different than zero when we condition firm decisions on the previous

year’s product characteristics. It also appears that there are no important efficiency gains occurring

during the decade but rather a small long term increase in cost of production perhaps driven by

factors associated to the long term increase in sales of larger and more powerful vehicles during the

1990s. Finally, costs are also increasing in the unobserved quality attribute, ξ. This may include

better performance measured as reliability (or torque for diesel vehicles) as well as cost associated

to setting up dealership networks for Asian newcomers.

In Figure 6 we report differences in marginal costs across brands relative to the Spanish

market leader, renault. Results are very reasonable, capturing the common perception of the

automobile market in Spain. German upscale brands audi, bmw, and mercedes, are among the

most expensive to produce. Chrysler (U.S. based) and Asian imports are quite competitive, with

Korean imports daewoo, hyundai, and kia, averaging a 26% relative cost advantage. European

manufacturers with lower unit costs of production than renault, include the Czech brand skoda

and the old Spanish brand seat, both of them acquired by volkswagen to sell streamlined versions

of their vehicles targeting lower income customers. Another interesting case of relatively low cost

of production is ford, which produces most of its smaller European models in a large plant located

in Spain. These results reassure us that our specification is reasonable and that our estimates will

be helpful in evaluating meaningful counterfactuals.

– 23 –

Table 2: Demand and Supply Estimates for Different Specifications

Model 1 Model 2 “BLP”

Coefficient Rob. SE Coefficient Rob. SE Coefficient Rob. SE

Standard Dev. (σ)

hpw 1.6258 (0.2060)∗∗∗ 0.6576 (0.3736)∗ 3.9037 (0.3717)∗∗∗

kpe 1.8003 (0.0908)∗∗∗ 1.9627 (0.0665)∗∗∗ 2.1083 (0.2665)∗∗∗

diesel 6.8429 (1.2848)∗∗∗ 7.1065 (1.4239)∗∗∗ 7.0359 (0.3987)∗∗∗

constant 3.8922 (0.8962)∗∗∗ 5.6884 (0.7858)∗∗∗ 2.1011 (0.3579)∗∗∗

Interactions (Π)

price/income −4.3618 (0.3875)∗∗∗ −4.4068 (0.3404)∗∗∗ −2.1563 (0.1764)∗∗∗

Mean Utility (β)

hpw −0.0770 (0.3687) 0.5972 (0.2105)∗∗∗ 3.9996 (1.3082)∗∗∗

size 5.0385 (0.3084)∗∗∗ 5.9951 (0.3246)∗∗∗ 5.7068 (1.0388)∗∗∗

diesel −9.6992 (0.2064)∗∗∗ −9.5499 (0.2092)∗∗∗ −10.6753 (0.8381)∗∗∗

diesel× trend 0.5579 (0.0416)∗∗∗ 0.6271 (0.0420)∗∗∗ 0.5016 (0.0386)∗∗∗

non-eu −1.4706 (0.1187)∗∗∗ −1.4944 (0.1204)∗∗∗ −1.1276 (0.1713)∗∗∗

seat 0.0491 (0.1756) 0.0542 (0.1758) 0.8839 (0.2791)∗∗∗

compact 1.2363 (0.1364)∗∗∗ 1.1549 (0.1372)∗∗∗ −0.4341 (0.3178)sedan 2.4353 (0.1314)∗∗∗ 2.3257 (0.1329)∗∗∗ −0.2111 (0.3842)luxury 4.7115 (0.1914)∗∗∗ 4.6207 (0.1933)∗∗∗ −0.2913 (0.4702)minivan 2.1268 (0.1629)∗∗∗ 2.0407 (0.1632)∗∗∗ −0.3776 (0.4149)constant −13.1307 (0.1663)∗∗∗ −16.7843 (0.1684)∗∗∗ −15.5799 (1.2450)∗∗∗

trend 0.0766 (0.0293)∗∗∗ 0.0897 (0.0298)∗∗∗ 0.0087 (0.0214)

Cost (γ)

ln(hpw× steel price) 0.2138 (0.0308)∗∗∗ 0.2502 (0.0317)∗∗∗ 0.8121 (0.0657)∗∗∗

ln(size× steel price) 1.0090 (0.0956)∗∗∗ 1.1696 (0.0900)∗∗∗ 2.5370 (0.1678)∗∗∗

ln(c90) 0.0211 (0.0130) 0.0308 (0.0121)∗∗ 0.2727 (0.0711)∗∗∗

unobserved demand (ξ) 0.0711 (0.0097)∗∗∗ 0.0585 (0.0076)∗∗∗

constant 0.7123 (0.0158)∗∗∗ 0.8375 (0.0148)∗∗∗ −0.0857 (0.1528)trend 0.0159 (0.0018)∗∗∗ 0.0129 (0.0017)∗∗∗ −0.0011 (0.0038)diesel 0.8333 (0.0132)∗∗∗ 0.7073 (0.0126)∗∗∗ 0.3662 (0.0354)∗∗∗

Elasticity Statistics:

- Average 5.7 5.9 3.1- Maximum 18.4 18.4 10.6- Minimum 2.9 3.1 1.4

Margin Statistics (%)

- Average 20.3 19.7 37.7- Maximum 38.1 36.1 74.7- Minimum 4.9 5.0 10.0

Estimation Statistics

Number of observations 1,740 1,740 1,740Simulated agents per year 5,000 5,000 5,000Information Set Xt−1,Wt−1, ξt−1 Xt,Wt, ξt N/AJ-Statistic 52.9 42.6 151.9

Robust standard errors in parentheses. Significant estimates with p-values less than 0.1, 0.05, and 0.01 are identifiedwith ∗, ∗∗, and ∗∗∗, respectively. Cost fixed effects for brand and segment not reported. b Estimates based onprojecting the estimated values of the demand unobservable ξ on other demand characteristics, including segmentfixed effects and a time trend. “Margin” defined as 100 × p−c

pwhere price excludes import tariffs, if applicable.

Equilibrium prices account for year-specific ownership structure as reported in Appendix A (Table A.1).

– 24 –

Figure 6: Production Costs Differences Across Brands(Reference: Renault)

-30%

-25%

-20%

-15%

-10%

-5%

0%

5%

10%

15%

20%

25%

30%

Alfa

Rom

eo

Aud

i

BM

W

Chr

ysle

r

Citr

oen

Dae

woo

Fiat

Ford

Hon

da

Hyu

ndai

Kia

Lanc

ia

Maz

da

Mer

cede

s

Mits

ubis

hi

Nis

san

Ope

l

Peug

eut

Rov

er

Saab

Seat

Skod

a

Suzu

ki

Toyo

ta

Vol

ksw

agen

Vol

vo

As for demand, Table 2 shows that it is downward slopping and always elastic. Both

specifications of full model generate similar price elasticities and price-cost margins. Using the

first specification as the benchmark, we estimate an average estimated price elasticity around 5.7

implying an average of 20% margin for the Spanish automobile industry during the 1990s. There is

however substantial heterogeneity, with margins as low as 4.9% and as high as 38.1 percent. This

wide range of margins are due to heterogeneous valuation of cars’ characteristics at a moment in

time, the evolution of preferences over time, and the changing product offering over the decade.23

Figure 7 shows that average estimated margins, both of gasoline and diesel vehicles, remain quite

stable, only decreasing very slightly during the middle 1990s for gasoline models. In the case of

diesel vehicles margins are relatively large at the beginning of the decade but they fall as the decade

proceeds due to increased imitation and competition in the segment. At the same time, consumer

23Although ignoring the distinction between diesel and gasoline models, Moral and Jaumandreu (2007) show thatdemand elasticities are smaller but also very heterogeneous across market segments and product life cycle.

– 25 –

preferences towards diesels are improving (βDiesel × Trend > 0) so the net effect is to stabilize diesel

margins over the latter half of the decade. For both engine types, the dispersion of margins grows

during the last three years of the sample as a growing economy increases dispersion of consumer

incomes.

Figure 7: Evolution of Estimated Price-Cost Margins (Model 1)

0

5

10

15

20

25

30

35

40

Mar

gin

(PC

M)

1992 1993 1994 1995 1996 1997 1998 1999 2000excludes outside values

(a) Diesel Vehicles

0

5

10

15

20

25

30

35

40

Mar

gin

(PC

M)

1992 1993 1994 1995 1996 1997 1998 1999 2000

excludes outside values

(b) Gasoline Vehicles

Estimates of Table 2 show that after we control for price and fuel efficiency, drivers prefer

larger cars (positive coefficient for size) even after controlling for segment. The average consumer is

indifferent about performance (insignificant coefficient for hpw in mean utility) but, not surprisingly

there is a great deal of heterogeneity as captured through the significant random coefficient for

hpw. The negative and significant sign of non-eu is an empirical regularity in the international

trade literature and is commonly referred to as the “home bias” effect.24 Since our focus is on a

specific industry rather than a set of bilateral trade flows across many sectors, we can provide a

more detailed interpretation. At this time, Asian imports were first sold in the European market

and were considered low quality, fuel-efficient alternatives to European vehicles but they lacked

both brand recognition as well as a widespread network of dealerships for maintenance. Thus, the

negative sign of non-eu is not surprising.

The large and significant value for kpe as a random coefficient indicates that Spanish drivers

are both concerned with fuel efficiency on average while their heterogenous tastes towards fuel

efficiency leads to different substitution patterns between fuel-efficient and fuel inefficient vehicles.

The results also indicate that some drivers strictly prefer diesel vehicles while on average diesels

are relatively unpopular after controlling for their primary competitive advantage, fuel efficiency.

We have thus far appealed to intuitive arguments to justify why we expect that observed

and unobserved product characteristics are likely to be correlated. Now that we have estimated

the model and recovered the unobserved quality index ξ, we can corroborate our intuition. Table 3

24See Cosar, Grieco, Li and Tintelnot (2015) for estimates of cross-country home bias in the automobile industry.

– 26 –

Table 3: Are Product Characteristics Correlated?

size hpw ξ

hpw 1.0000-

size 0.3921 1.0000(0.0187) -

ξ 0.7171 0.3109 1.0000(0.0128) (0.0199) -

Notes: Estimated unobserved product characteristics (ξ) based onprevious year characteristics (i.e., Model 1).Standard errors reportedin parentheses.

presents the correlations between the observable product characteristics included in the BLP esti-

mation and the estimated unobserved product characteristic ξ implied by our alternative estimation

approach. The reported results provide clear evidence that the observed and unobserved product

characteristics are indeed very much correlated – consistent with the results of Petrin and Seo

(2016).

The natural question then is whether these correlations are quantitatively important. In

Table 2 we juxtapose estimates from our full model with estimation results when we assume that

product characteristics are exogenous (model “BLP”). In an attempt to keep the specifications as

close as possible to each other, we maintained the same set of random coefficients and demographic

interactions as well as the same simulated agents. We also included the post-gmm regressors from

our full model into the gmm BLP estimation. Our only departure from Berry et al. (1995) pertains

to the instruments which we found uninformative for our data. Instead we employed “differentiation

IVs” introduced by Gandhi and Houde (2015). The idea is similar to the concentration instruments

proposed in Berry et al. (1995) but instead of simple sums or averages (e.g., sum of hpw for

products not owned by the producer of product j in period t), we include summary statistics for

the distribution of distances in product space (e.g., average distance in hpw of products not owned

by the producer of product j and far away from product j in hpw space).

The comparison reveals that assuming orthogonality produces significant estimation biases.

The most striking feature between the estimation approaches is the difference in the price coefficient

where we find a much smaller value when we impose product exogeneity, leading to lower estimated

price elasticities and higher estimated markups. We also find larger random coefficients for hpw

and kpe but a smaller random coefficient for the constant. The model’s overall fit is also poor – a

common result among BLP models – reflected in a J-statistic of 151.9 which is substantially greater

than the 95% critical value. All of this leads us to reject the assumption of product characteristic

exogeneity.

– 27 –

6 Emissions Policy as Industry Protection

In this section we present evidence that emissions standards did indeed drive the rise of diesels

in Europe. Said differently, we show that had the EU imposed more rigorous NOx emissions

standards the diffusion of diesels would have been much smaller. The trade implication of such

policies are significant – a near doubling the import share from 11.8 to 19 percent. By not adopting

such damaging policies, whether inadvertently or not, European policymakers implicitly helped

European manufacturers enhance their dominance in the domestic market. To our knowledge this

is the first time a structural equilibrium model of oligopoly industry competition in differentiated

products has been used to evaluate trade frictions, in particular the trade protective impact of

environmental regulation.

We model a change in emissions policy as an in increase in marginal cost applicable only to

diesel vehicles. We think of this “retrofitting” cost as the additional equipment required to make

the diesel fleet compliant with the new standard. We assume that all diesel models require the

same cost. The task then is to identify a “realistic” cost to retrofit an auto maker’s diesel fleet to

the new standard. For years, a technology to successfully capture NOx emissions at the tailpipe

simply did not exist. When it finally became available, in the late 2000s, it was still very expensive.

By the EPA’s own estimates in 2010, diesel engines could be retrofitted to comply with both EPA

and California NOx emission standards by means of a Lean NOx Catalyst at an estimated cost of

between $6,500 to $10,000 per vehicle. Lean NOx catalysts use diesel fuel injected into the exhaust

stream to create a catalytic reaction and reduce pollution. However, these catalysts still require

specific exhaust temperatures for appropriate NOx emission control performance, and on average

they reduce emissions up to a maximum of 40%. German manufacturers bmw and mercedes

were certified to be sold in all 50 states of the U.S. in 2009 only after equipping their new vehicles

with a Selective Catalytic Reduction System that injects a reluctant (a urea-based solution) into

the exhaust stream where it reacts with a catalyst to convert NOx emissions to nitrogen gas and

oxygen. This system is more effective, reducing NOx emissions up to 75% but the EPA estimated

that its cost ranged between $10,000 and $20,000 per vehicle in 2010.25

Computing the equilibrium of a counterfactual environmental policy reintroduces issues

about multiple equilibria as firms could potentially adjust both product characteristics and prices

in response to an alternate regulation. Rather than computing all potential equilibria and evaluating

the consequences of each, we take a simpler approach and keep the product characteristics fixed

25On retrofitting costs see Diesel Retrofit Devices. EPA’s National Clean Diesel Campaign, 2013. http://www.epa.

gov/cleandiesel/technologies/retrofits.htm as well as our summary in Appendix C.

– 28 –

while allowing firms to re-optimize their prices.26 While there is no guarantee of uniqueness in the

pricing game either, we’ve found little evidence of multiple equilibria as solving the pricing game

from different initial conditions leads to the same equilibrium.

We believe this simpler approach is indeed a good first-order approximation to the short-

term (i.e., the 1990s) impacts of alternative emissions and fuel taxation policies. Evidence suggests

that product innovations in this industry are slow as redesigning a car’s size, engine, drive-chain, etc.

is costly and time-consuming – an issue consistent with our estimation. This is likely particularly

true for changes in the characteristic which diesels hold a competitive advantage, fuel efficiency,

since designing and redesigning engines are both extremely expensive and time consuming.27 This

would suggest that our results are best thought of as a realistic approximation for the short-term

(though still denominated in years) effects of emissions policy, while the longer term (i.e., 10-20

years) are unclear.

Our final task is to choose a benchmark set of estimation results which we can then use to

assess the implications of different emissions policies. Given the above results we choose the model

in which firms form expectations using the previous year’s product set as our benchmark (i.e.,

Model 1). Our rationale is that this information set demands far less knowledge and foresight from

the firms. Since both specifications of the full model generate similar demand and cost estimates,

our selection of this model, fortunately, has little quantitative or qualitative implication on our

results.

6.1 Consumer Response to a More Rigorous NOx Policy

In Figure 8 we present the consumer response as we vary the additional cost of retrofitting expense

(x-axis) required to make diesels compliant with a more rigorous NOx emissions policy. The shaded

area highlights the limits of the estimated EPA retrofitting costs corrected for the exchange rate

and inflation.

Figure 8 shows that the increase in production costs required to comply with the more

rigorous NOx environmental regulation leads consumers to substitute away from diesel vehicles

for even small retrofitting costs. At the lower bound for the EPA estimate, a retrofitting cost of

26Specifically, one could identify all potential equilibria using a homotopy algorithm as in Besanko, Doraszelski,Kryukov and Satterthwaite (2010). Although difficult, one could presumably evaluate the likely impact of apolicy change by assessing commonalities between the equilibria (e.g., change in diesel market share and profits).A complicating factor is that an alternative NOx emissions policy would have likely impacted the evolution ofconsumer preferences for the diesel over the decade; a fact currently captured in the positive and significantestimate for the diesel trend variable βDiesel × Trend.

27Busser and Sadoi (2004, Footnote 2) document that since demand was small in their countries of origin, Asianmanufacturers such as Toyota acquired engines from other European firms as a less costly way to satisfy Europeandemand rather than investing in the development of diesel engines from scratch.

– 29 –

Figure 8: Market Shares, Quantity Sold, and Retrofitting Costs

EU: Gas

Non-EU: Gas

EU: DieselNon-EU: Diesel0

10

20

30

40

50

60

70

80

Mar

ket S

hare

(%

)

0 5 10 15 20 25

Retrofitting Expense (Thousands of Euro)

e3,300, we already see a significantly negative impact on the popularity of diesel vehicles as diesel

market share falls by 12 percent. At the upper bound, a retrofitting cost of e6,600, the market

share of European diesel vehicles is nearly cut in half. For a retrofitting cost of e12,000 the market

share of European diesels returns to the level observed at the beginning of the sample and below

the share of gasoline imports, who grow monotonically with the retrofitting costs although the sales

of European gasoline models grows much faster.

A detailed analysis of market shares of the different manufacturers reveals that the only

clear beneficiaries of an alternative stringent European NOx emission policy would be foreign

automakers. As producers of inexpensive, fuel-efficient gasoline vehicles, the foreign auto makers

benefit as consumers substitute away from the expensive diesel engines and towards gasoline.

Although the composition of sales changes with retrofitting costs, most European manufacturers

maintain a significant share of the market. That is not the case for the two European diesel

leaders psa and volkswagen. Both of them are also the largest producers of diesel vehicles in

Europe and thus, having to face these large retrofitting costs erode their competitiveness and their

market shares. Lastly, it is important to note that the consumer’s general preference for domestic

automakers (βNonEU < 0) limits their willingness to choose foreign imports. Had this home bias

not existed, the shift from domestic to foreign cars would have been much larger.

– 30 –

6.2 Impact to Firms

The emphasis on market shares hides aggregate industry effects as more stringent emissions policy

leads to higher retail prices and lower aggregate sales (Figure 9). The mechanism driving this

decline is straightforward. While the model does allow for rich substitution patterns for consumers

to switch to an alternative new car if the price of their first choice increases, the fact the alternative

emissions policy increases the costs for a large number of cars simultaneously leads firms to increase

price for both diesel and gasoline vehicles.

Consumers respond by leaving the market (panel b), presumably to purchase a used vehicle

or to hold-off consumption altogether.28 The results for firms, particularly European firms, are

equally negative (panel a). volkswagen and the psa group, the heaviest adopters of diesels,

are the biggest losers though most European firms experience a material reduction in profits. In

comparison, Asian firms which had invested little in developing diesel products see their profits

increase as consumers switch to their fuel-efficient gasoline models.

Figure 9: Impact of Alternative Emissions Policies on Firms

EU: Gas

Non-EU: Gas

EU: DieselNon-EU: Diesel0

500

1000

1500

2000

Pro

fit (

M)

0 5 10 15 20 25

Retrofitting Expense (Thousands of Euro)

(a) Profit (eMillions)

800

1000

1200

1400

1600

Qua

ntity

Sol

d (0

00)

0 5 10 15 20

Retrofitting Expense (Thousands of Euro)

(b) Quantity (000)

To quantify the aggregate impact of diesels, Table 4 reports the extreme case where an

alternative emissions policy results in the elimination of diesels from the marketplace. For simplicity

we just report results for year 2000 but other years yield similar results. While we admit this is

a limiting case, it is worthwhile to remind the reader that this is indeed what happened in the

United States market after CAAA implementation in the mid 1990s. Relative to the benchmark,

the market shrinks considerably as consumers move to the outside good (e.g., used car market).

European automakers absorb the most of this impact as profits fall e 2.2 billion, or 64 percent.

Non-European, primarily Asian, auto makers, however, are better off as their market share jumps

28By holding the value of the outside option (i.e., not buying a new car) fixed, we are assuming the average priceand quality of a used car does not change with the regulation.

– 31 –

from 11.8% to 19.6% resulting in more market power and an increase retail prices, margins, and

sales of their gasoline models following the disappearance of fuel-efficient diesel vehicles from the

market. On net, profits for these firms increase e 21 million, or 5.5 percent.

Table 4: Value of the Diesel

Scenario Models Price Quantity Margin Share Profit

Benchmarkeu: diesel 75 16.19 695.37 18.68 50.95 1,961.00eu: gasoline 84 14.93 508.70 21.09 37.28 1,434.37non-eu: diesel 20 17.20 36.97 14.84 2.71 83.26non-eu: gasoline 50 13.66 123.65 21.18 9.06 291.05

Equilibrium without Dieselseu: gasoline 84 21.11 412.58 16.95 80.40 1,236.29non-eu: gasoline 50 18.03 100.58 25.83 19.60 394.89

Results based on year 2000 equilibrium. “Price” is the sales-weighted average price faced by consumers (inthousands of 1994 Euros), including tariffs. “Quantity” is measured in millions of cars. “Profit” is measured inthe equivalent of millions of 1994 Euro. “Margin” and “Share” are reported as percentages. “Margins” includeimport duties paid by consumers.

6.3 Import Tariff Equivalence of Environmental Regulation

We have so far shown that diesel vehicles were a popular choice among consumers; generating

substantial profits for European auto makers. Reducing the popularity of these vehicles, presumably

from an EPA-like emissions policy, would have resulted in a substantial reduction in profits for

these firms while nearly doubling the market share of imports. In this section we use the structural

model to measure the tariff-equivalence of European authorities’ targeting of green house emissions.

Whether this emission policy was designed explicitly to promote sales of domestically produced

diesel vehicles is inconsequential. In practice, targeting CO2 rather than NOx generated precisely

that result.

In Figure 10 we plot the import tariff required to generate the import share we observe

(e.g., 11.8% in 2000) for each level of retrofitting cost. We interpret each point as the import

tariff-equivalence of the diesel-friendly emissions policy employed by EU regulators. For simplicity,

we again restrict the current discussion to the year 2000 and show in Appendix F.1 (Table F.2) the

results are similar for other years as well.

We find that the implicit tariff from by our “Baseline” estimation is significant, ranging

from 13.4% when retrofitting costs are on the lower bound of the EPA estimates to 16.4% at the

upper bound. For reference, the official import tariff facing Asian imports was 10.3% in 2000 so

the diesel-friendly EU emissions policy amounted to 30-60 percent increase on the official rate, a

significant effect. In the extreme case, a retrofitting cost which effectively eliminates the diesel

– 32 –

Figure 10: Import Tariff Equivalent of EU Emissions Policy

BaselineNo Trend

10

12

14

16

18

20

22

24

26

28

30

Impl

icit

Tar

iff (

%)

0 5 10 15 20 25

Retrofitting Expense (Thousands of Euro)

corresponds to a 26.5% import tariff. In other words, if the costs of modifying the diesel engine to

meet stronger NOx emissions standards are very large, the protective effect of adopting an emissions

policy with a weak NOx emissions standard is equivalent to imposing a tariff two-and-a-half times

the official rate.

An advantage of our data set is that it covers a period of massive adoption of the diesel

technology. In the estimation we allowed for consumer preferences towards diesels to change over

the decade due to potentially learning or unobserved improvements in these vehicles. In the analysis

thus far, we’ve held consumer preferences as fixed and only allowed substitution through changes in

vehicle price due to retrofitting costs. We argue that the implementation of an alternative emissions

policy early in the decade would not only affect immediate diesel sales through the changes in price

but also future sales, particularly if one believes the increasing favorability of consumers is due to

learning about the new technology. One would expect the EU’s pro-diesel policy to increase the

competitive advantage of diesels by buying time for consumers to learn about them. Conversely,

an alternative emissions policy would limit consumer adoption due to increases in price, leave

preferences towards diesels unchanged, and require a larger tariff to defend domestic auto makers.

In Figure 10 we confirm this hypothesis by presenting results where we allow for the change

in emissions policy to impact diesel demand through this second channel – i.e., through consumer

learning. We do this in the model by setting the diesel trend equal to zero (βDiesel × Trend = 0).

The underlying assumption is that a positive estimate for the diesel trend in demand is due solely

– 33 –

to increased consumer awareness of this next generation diesel technology. We then solve for the

implicit tariff as in the “Baseline” experiment; finding implicit tariff increases from 13 to 24%

at the lower bound of the EPA estimate and from 16 to 25% at the upper bound. Further, the

“No Trend” curve is flatter suggesting that small retrofitting costs effectively eliminate the diesel

segment. While we view this experiment as aggressive, it does reveal the quantitative importance

of consumer learning in the evaluation of any policy that promotes diffusion of a new technology.

For us, it also reveals that our “Baseline” estimates likely understate the trade implications.

Further refinement of the estimated implicit tariff depends crucially on pinning down a

“realistic” retrofitting cost which is complicated since the reference EPA estimates are based on

technology developed much later (2010) and one can imagine that European auto makers may have

been able to develop a less expensive technology to protect their investment in diesels. The recent

Volkswagen scandal suggests, however, that the costs of modifying these engines are indeed large

since the company chose to incur severe financial penalties rather than meet the stricter EPA’s NOx

thresholds. We take this as further evidence that a conservative estimate for the tariff-equivalent of

observed EU emissions policy is between 13.4 and 16.4 percent, though we note that these values

likely understates the actual effect provided one believes that the growth in diesels was due at least

in part to consumer learning. Regardless, it is clear the emissions policy employed by European

regulators favored domestic auto makers as a quantitatively significant de facto non-tariff trade

policy during the 1990s.

6.4 Impact to Consumers

Thus far our analysis has focused on the impact of emissions regulation on firms in the automobile

industry, particularly domestic and foreign firms. In this section we pivot to focus on consumers.

In Figure 11 we show the amount of money required to compensate the average consumer in

year 2000 under more rigorous emissions policies. The solid line indicates the average consumer

requires e150-250 in compensation for the retrofitting costs within the range estimated by the

EPA. The dashed line shows that consumers are further hurt (i.e., require more compensation)

when government imposes the “Baseline” import tariffs from Figure 10.

– 34 –

Figure 11: Impact of Alternative EmissionsPolicies on Consumers

Baseline

Tariff

0

50

100

150

200

250

300

350

400

450

500

550

600

650C

ompe

nsat

ing

Var

iatio

n P

er C

apita

(E

uros

)

0 5 10 15 20 25

Retrofitting Expense (Thousands of Euro)

While our welfare analysis does not account for the reduction in negative health externalities

due to more rigorous NOx standards, it does provide an interesting insight into non-tariff trade

policies more broadly. Recall that so far we have shown that the emissions policy employed by EU

regulators had the effect of promoting a domestic innovation. Further, we show in Figure 11 that

this policy is unambiguously welfare improving if we assume health externalities are negligible.

Putting these points together indicates that non-tariff trade policies which promote domestic

innovations (or the adoption of products by domestic consumers) can not only be an effective

tool to influence consumption towards domestic products but they may also improve consumer

welfare. This stands in stark contrast with tariffs which influence consumption by distorting price,

leading to less consumer surplus – a fact also illustrated in Figure 11.29

7 Robustness

In Section 5 we demonstrated that assuming product characteristic exogeneity led to steeper

estimated demand curves and greater margins. The natural question is whether these differences

materially impact our conclusions as to the effects of emissions policy on the European auto

industry. In Figure 12 we compare the total profits of European firms (left panel) and the share of

29There is a large and growing empirical literature documenting the negative effects of tariffs on consumer welfare.See Ruhl (2008) for a review.

– 35 –

total profits contributed by diesels (right panel) as we vary the retrofitting costs across our baseline

estimation (“Model”) and the alternative BLP estimates (“BLP”) from Table 2.

Figure 12: Value of Diesels by Estimation Approach

Model

BLP

0

500

1000

1500

2000

2500

3000

3500

4000

4500

5000

5500

6000

6500

7000

Pro

fit (

M E

uro)

0 5 10 15 20 25

Retrofitting Expense (Thousands of Euro)

(a) Total Profits

Model

BLP

0

5

10

15

20

25

30

35

40

45

50

55

60

Per

cent

of P

rofit

from

Die

sels

0 5 10 15 20 25

Retrofitting Expense (Thousands of Euro)

(b) Share of Profits from Diesels

The steeper demand estimates from the BLP estimation leads to larger margins and greater

profits for all products, not just diesels. We see this aggregate effect in a shifting up of the

profit curves when moving from our baseline specification to the BLP estimates. The right panel

illustrates that the shift is not uniform as diesels in our baseline model decrease in importance at

a faster rate than in the BLP specification as the larger estimate for the diesel random coefficient

in the latter specification makes consumers more willing to stay with the engine type despite

increases in price. The lower estimated price elasticities in the BLP estimation also enables auto

makers to pass-through increases in marginal cost to consumers leading to a more linear relationship

between retrofitting costs and the importance of diesels, measured in total profits or in market share.

We interpret these results as evidence that assuming product characteristic exogeneity biases the

estimated value of diesel vehicles upwards, thereby overstating the effects of pro-diesel emissions

policy on automakers and consumers alike.

Interestingly, this bias largely disappears, particularly for low estimates of the retrofitting

costs, when we compare the implicit import tariffs from each demand estimation (Figure 13). Why

does the estimation matter when analyzing the importance of diesels but not when we compare the

implicit import tariff implied by these vehicles? The answer lies in the firms’ first-order conditions

for price:

p = [1 + τ ]× [mc + ∆−1(p, x, ξ; θ)sj(p, x, ξ; θ)︸ ︷︷ ︸markups

] , (18)

The different estimation approaches impact our conclusions to the overall value of diesels

through the markups via the estimated product elasticities, e.g., see Panel (a) of Figure 12. Given

– 36 –

Figure 13: Implicit Import Tariff by Estimation Approach

Baseline

BLP

10

12

14

16

18

20

22

24

26

28

30

32

34

Impl

icit

Tar

iff (

%)

0 5 10 15 20 25

Retrofitting Expense (Thousands of Euro)

that we observe price and use (18) to back-out marginal costs consistent with profit-maximization,

the estimated price elasticities help the researcher split observed prices into marginal costs and

markups. Therefore, under the BLP approach we find that consumers are less price-sensitive than

under our baseline model which, in turn, leads us to allocate more of the observed prices to markups

and greater profits for all cars, including diesels.

Recall that in Figure 13 we vary the import tariff τ to move the import share for each

retrofitting cost back to the value observed in the data. Importantly, the import tariff τ increases

marginal costs and markups proportionately so the relative size of the two matters less. Conse-

quently, the net impact of using the simpler BLP approach depends on the complex substitution

patterns embedded in the markup function, ∆−1(p, x, ξ; θ)sj(p, x, ξ; θ), which in this case appears

to be small particularly in a neighborhood of the equilibrium observed in the data.

We use the above evidence to form two conclusions. The first is cautionary as Figure 12

clearly shows that a researcher who uses a discrete choice model to estimate markups (e.g., quan-

tifying the value of a good) should be wary of estimates based on product exogeneity as such an

assumption will significantly bias her results upwards. The second is more optimistic as researchers

interested in economic mechanisms in which the estimated markups are not of themselves important

may be able to use the much simpler BLP approach to generate reasonable results. Proving

this distinction more formally is an interesting and, given the prevalence of BLP -type models,

potentially important area of future research.

– 37 –

8 Concluding Remarks

The goal in this paper was to evaluate the role of vehicle emissions policy in the growth of the

diesel segment in the European automobile industry. To do so we estimated a structural oligopoly

model of differentiated products where we allowed for correlation between observed and unobserved

product characteristics, finding the two are indeed correlated. Our estimation allowed for significant

heterogeneity of preferences, finding that consumers not only favor fuel efficiency and car size but

also that their perception of diesels improved dramatically in the decade following the introduction

of these next generation engines. We also find that widespread imitation of the tdi by European

auto makers due to the generality of the technology enabled domestic firms to generate substantial

profits from the technology.

Perhaps the most novel result of our paper is to show that seemingly non-trade policies

such as environmental standards can have important and quantitatively significant trade effects.

Regardless of whether the European pro-greenhouse emission policy was intended to favor the

sales of domestically produced diesel vehicles or not, we show that alternative NOx reduction

policies would have effectively halted the commercial success of diesel vehicles. Further, the diesel-

friendly emissions policy favored domestic firms equivalent to an implicit import tariff ranging

from 13.4% to 16.4%, or approximately a 30 to 60% increase over the official tariff employed by

the European Union at the time. Moreover, we show this result is robust across a variety of

assumptions, including the estimation approach. This is, to the best of our knowledge, the first use

of a structural equilibrium model of demand and industry oligopoly competition to evaluate the

trade effects of a non-tariff policy. Our results illustrate that in an increasingly global economy,

governments can effectively construct non-trade oriented national policies, including environmental

regulations, to protect domestic industries when traditional trade policies are no longer available.

We further show that, in contrast to tariffs, such a policy may be welfare improving.

While our modeling choices are sufficient to address the objectives in this paper – balancing

a feasible extension of the BLP framework while meeting the institutional details of our application

– we view this paper as a step towards developing a more realistic empirical model of the automobile

industry by providing useful insights into the quantitative implications of attribute choices made

by firms.30

30Crawford (2012) presents a recent overview of the challenges of fully addressing the endogeneity of product attributesin discrete choice models of demand.

– 38 –

References

Aw, B. Y., M. J. Roberts, and D. Y. Xu (2011) “R&D Investment, Exporting, and ProductivityDynamics,” American Economic Review, Vol. 101, No. 4, pp. 1312–44, June.

Bagwell, K. and R. W. Staiger (2001) “Domestic Policies, National Sovereignty and InternationalEconomics Institutions,” Quarterly Journal of Economics, Vol. 116, pp. 519–562.

Bajari, P., C. L. Benkard, and J. Levin (2007) “Estimating Dynamic Models of Imperfect Compe-tition,” Econometrica, Vol. 75, No. 5, pp. 1331–1370.

Balcer, Y. and S. A. Lippman (1984) “Technological Expectations and Adoption of ImprovedTechnology,” Journal of Economic Theory, Vol. 34, pp. 292–318.

Barrett, G. F. and S. G. Donald (2003) “Consistent Test for Stochastic Dominance,” Econometrica,Vol. 71, pp. 71–104.

Berry, S., J. Levinsohn, and A. Pakes (1995) “Automobile Prices in Market Equilibrium,” Econo-metrica, Vol. 63, pp. 841–890.

(1999) “Voluntary Export Restraints on Automobiles: Evaluating Strategic Trade Policy,”American Economic Review, Vol. 89, pp. 400–430.

Besanko, D., U. Doraszelski, Y. Kryukov, and M. Satterthwaite (2010) “Learning-by-Doing, Orga-nizational Forgetting, and Industry Dynamics,” Econometrica, Vol. 78, No. 2, pp. 453–508.

Bhagwati, J. and V. K. Ramaswami (1963) “Domestic Distortions, Tariffs, and the Theory ofOptimum Subsidy,” Journal of Political Economy, Vol. 71, pp. 44–50.

Bresnahan, T. F. (2010) “General Purpose Technology,” in B. H. Hall and N. Rosemberg eds.Handbook of the Economics of Innovation, Vol. II, New York, NY: North-Holland.

Busser, R. and Y. Sadoi (2004) “Japanese Automobile Makers in Europe and the Organizationof the Supply System,” in R. Busser and Y. Sadoi eds. Production Networks in Asia andEurope. Skill Formation and Technology Transfer in the Automobile Industry, New York,NY: Taylor & Francis.

Canis, B. (2012) “Why Some Fuel-Efficient Vehicles Are Not Sold Domestically,” report to congress,Congressional Research Service.

Chamberlain, G. (1987) “Asymptotic Efficiency in Estimation with Conditional Moment Restric-tions,” Journal of Econometrics, Vol. 34, pp. 305–334.

Cosar, K., P. Grieco, S. Li, and F. Tintelnot (2015) “What Drives Home Market Advantage?”.manuscript.

Crawford, G. S. (2012) “Endogenous Product Choice: A Progress Report,” International Journalof Industrial Organization, Vol. 30, pp. 315–320.

Deardorff, A. V. (1996) “An Economist’s Overview of the World Trade Organization,” WorkingPaper 388, Research Seminar in International Economics, University of Michigan.

– 39 –

Dube, J.-P., J. T. Fox, and C.-L. Su (2012) “Improving the Numerical Performance of Static andDynamic Aggregate Discrete Choice Random Coefficients Demand Estimation,” Economet-rica, Vol. 80, pp. 2231–2267.

Ederington, J. and J. Minier (2003) “Is Environmental Policy a Secondary Trade Barrier? AnEmpirical Analysis,” Canadian Journal of Economics, Vol. 36, pp. 1371–154.

Edmond, C., V. Midrigan, and D. Y. Xu (2015) “Competition, Markups, and the Gains fromInternational Trade,” American Economic Review, Vol. 105, pp. 3183–3221.

Feenstra, R. C. (1988) “Quality Change Under Trade Restraints in Japanese Autos,” QuarterlyJournal of Economics, Vol. 103, pp. 131–146.

Gandhi, A. and J.-F. Houde (2015) “Measuring Substitution Patterns in Differentiated ProductsIndustries.” manuscript.

Grigolon, L., M. Reynaert, and F. Verboven (2015) “Consumer Valuation of Fuel Costs and theEffectiveness of Tax Policy: Evidence from the European Car Market.” manuscript.

Hansen, L. and K. J. Singleton (1982) “Generalized Instrumental Variables Estimation of NonlinearRational Expectations Models,” Econometrica, Vol. 50, No. 5, pp. 1269–86.

Manuelli, R. and A. Seshadri (2014) “Frictionless Technology Diffusion: The Case of Tractors,”American Economic Review, Vol. 104, No. 4, pp. 1368–1391.

Moral, M. J. and J. Jaumandreu (2007) “Automobile Demand, Model Cycle and Age Effects,”Spanish Economic Review, Vol. 9, pp. 193–218.

Nevo, A. (2000) “Mergers with differentiated Products: The Case of the Ready-to-Eat CerealIndustry,” RAND Journal of Economics, Vol. 31, pp. 395–421.

Newey, W. K. (1990) “Efficient Instrumental Variables Estimation of Nonlinear Models,” Econo-metrica, Vol. 58, pp. 809–837.

(1993) “Efficient Estimation of Models with Conditional Moment Restrictions,” in Hand-book of Statistics, Vol. 11: Elsevier, pp. 419–?454.

Pakes, A., J. Porter, K. Ho, and J. Ishii (2015) “Moment Inequalities and Their Application,”Econometrica, Vol. 83, No. 1, pp. 315–334.

Petrin, A. and B. Seo (2016) “Identification and Estimation of Discrete Choice Models whenObserved and Unobserved Product Characteristics are Correlated.” manuscript.

Ruhl, K. J. (2008) “The International Elasticity Puzzle.” manuscript.

Schumpeter, J. A. (1950) Capitalism, Socialism, and Democracy, New York, NY: Harper & Row,3rd edition.

Staiger, R. W. (1995) “A Theory of Gradual Trade Liberalization,” in A. V. Deardorff, J. A.Levinsohn, and R. M. Stern eds. New Directions in Trade Theory, Ann Arbor, MI: TheUniversity of Michigan Press.

– 40 –

Stewart, S. J. (2010) “Collection of PM Emission Data from In-Use Light-Duty Diesel and GasolineVehicles,”Technical report, British Columbia Clean Air Research Fund Final Report, PacificVehicle Testing Technology, Ltd.

Thurk, J. (2014) “The Effects of Coordinating National Patent Policies,” Mimeo, University ofNotre Dame.

Verboven, F. (2002) “Quality-Based Price Discrimination and Tax Incidence: Evidence from Gaso-line and Diesel Cars,” RAND Journal of Economics, Vol. 33, pp. 275–297.

– 41 –

Appendix

A Spanish Data Sources

To control for household income distribution a thousand individuals are sampled each year from

the Encuesta Continua de Presupuestos Familiares (Base 1987 for years 1992-1997 and Base 1997

for years 1998-2000) conducted by INE, the Spanish Statistical Agency.31 The outside option varies

significantly during the 1990s due to the important recession between 1992 and 1994 and the very

fast growth of the economy and population (immigration) in the second half of the decade. We

also use these consumer surveys to set the size of the outside option for each year in our sample.

Starting with 1992, they are: 0.92, 0.94, 0.93, 0.93, 0.93, 0.92, 0.91, 0.89, and 0.89, respectively.

Fuel prices were also obtained from INE. In real 1994 euro-equivalent denominations per

liter, these are 0.445, 0.488, 0.490, 0.493, 0.543, 0.560. 0.530, 0.565, and 0.695 for diesel and

0.580, 0.628, 0.655, 0.678, 0.706, 0.724, 0.702, 0.737, and 0.875 for gasoline, for years 1992 to 2000,

respectively. As for the Spanish steel prices used as instruments for the cost equations, they are

obtained from the 2001 edition of Iron and Steel Statistics – Data 1991-2000 published by the

European Commission (Table 8.1).

For the analysis of demand we build a data set using prices and vehicle characteristics

as reported by La guıa del comprador de coches, ed. Moredi, Madrid. We select the price and

characteristics of the mid-range version of each model, i.e., the most popular and commonly sold.

Demand estimation also makes use of segment dummies. Other than the luxury segment, which

also includes sporty cars, our car segments follow the “Euro Car Segment” definition described

in Section IV of “Case No. COMP/M.1406 - Hyundai/Kia.” Regulation (EEC) No. 4064/89:

Merger Procedure Article 6(1)(b) Decision. Brussels, 17 March 1999. CELEX Database Document

No. 399M1406.

Until Spain ended its accession to the European Union transition period in 1992, it was

allowed to charge import duties on European products. Similarly, import duties for non-European

products converged to European levels. European imports paid tax duty of 4.4% in 1992, and

nothing thereafter. Non-European manufacturers had to pay 14.4% and 10.3%, respectively.

Thus, for the estimation of the equilibrium random coefficient discrete choice model of Table 2

we distinguish between prices paid by consumers (p) and those chosen by manufacturers (pτ ) .

The other relevant factor that changes during the 1990s is the ownership structure of

automobile firms. During this decade fiat acquired alfa romeo and lancia; ford acquired

volvo; and gm acquired saab. bmw acquired rover in 1994 but sold it in May 2000 (with the

exception of the “Mini” brand) so these are treated as separate firms. Table A.1 describes the

ownership structure at the beginning and end of the decade.

31See http://www.ine.es/jaxi/menu.do?L=1&type=pcaxis&path=/t25/p458&file=inebase for a description ofthese databases in English.

– i –

Table A.1: Automobile Groups: 1992 vs. 2000

Year 1992 Year 2000

Firm Gasoline Diesel Owner Gasoline Diesel Owner

alfa romeo 5,038 64 alfa romeo 2,941 3,983 fiataudi 16,689 1,982 volkswagen 15,273 24,184 volkswagenbmw 17,855 1,906 bmw 13,683 15,838 bmwchrysler 1,243 – 5,941 2,389citroen 68,890 36,851 psa 46,420 111,694 psadaewoo – – 25,201 –fiat 35,677 5,733 fiat 30,557 17,967 fiatford 121,140 17,468 ford 55,268 57,013 fordhonda 4,805 – 8,782 1,072hyundai 2,704 – 30,150 3,590kia – – 9,778 1,387lancia 11,117 905 lancia 2,206 2,126 fiatmazda 3,064 – 2,205 1,480mercedes 9,352 4,129 mercedes 13,953 10,684 mercedesmitsubishi 3,041 – 3,660 1,013nissan 16,010 905 17,855 21,971opel 110,286 11,099 gm 66,488 75,418 gmpeugeot 61,323 35,494 psa 55,371 92,496 psarenault 147,907 27,448 renault 76,925 99,360 renaultrover 15,255 425 rover 10,173 8,491 roversaab 1,551 – saab 1,867 2,424 gmseat 85,773 11,787 volkswagen 58,072 109,447 volkswagenskoda 724 – skoda 5,003 10,385 volkswagensuzuki 2,058 – 3,250 486toyota 4,425 – 16,827 3,584volkswagen 50,561 5,471 volkswagen 47,125 50,296 volkswagenvolvo 10,179 – volvo 7,379 3,566 ford

Sales of vehicle by manufacturer and fuel type. “Owner” indicates the name of the automobilegroup with direct control on production and pricing. Those without a group are all non-European manufacturers and defined as non-eu in the analysis.

– ii –

Figure A.1: Sales by Firm and Type of Engine

0 25000 50000 75000 100000 125000 150000

SKODACHRYSLER

SAABSUZUKI

HYUNDAIMITSUBISMAZDATOYOTAHONDAVOLVO

ALFA ROMROVERLANCIANISSANBMWAUDI

MERCEDESVOLKSWAGEN

FIATOPELSEATFORD

RENAULTPEUGEOTCITROEN

Diesel Gasoline

(a) Year 1992

0 25000 50000 75000 100000 125000 150000

DAEWOOSUZUKI

MITSUBISHONDA

KIAMAZDALANCIA

CHRYSLERSAAB

VOLVOTOYOTAHYUNDAIALFA ROM

ROVERSKODA

MERCEDESBMWFIAT

NISSANAUDI

VOLKSWAGENFORDOPEL

PEUGEOTRENAULT

SEATCITROEN

Diesel Gasoline

(b) Year 2000

– iii –

Table A.2: Car Model Characteristics Across Engine Types

segment models share price c90 kpe size hpw

1992

Compact 31 35.79 10.96 5.33 32.07 74.34 3.98Sedan 39 22.31 14.26 5.69 30.27 80.10 4.26Luxury 39 5.77 24.01 6.49 25.75 87.07 4.84Minivan 4 0.32 17.28 6.93 24.21 81.66 3.79Small 28 35.82 7.98 4.68 35.00 62.51 3.65

All 141 100.00 11.40 5.25 32.33 72.15 3.97

2000

Compact 56 34.43 14.86 5.00 32.53 76.54 3.59Sedan 52 25.97 19.45 5.26 31.60 81.92 3.63Luxury 40 3.72 34.53 6.72 23.31 89.72 5.17Minivan 32 3.13 20.80 6.39 25.91 83.47 3.16Small 49 32.75 10.42 4.86 31.61 66.36 3.18

All 229 100.00 15.52 5.13 31.43 75.31 3.51

Notes: share is the market share as defined by automobiles sold. price is denominated in the equivalentof thousands of 1994 Euros and includes value added taxes and import tariffs. kpe is the distance,measured in kilometers, traveled per euro of fuel. size is length×width measured in square feet. hpw isthe performance ratio of horsepower per hundred pounds of weight.

– iv –

Figure A.2: Diesel Adoption Rates

0.0

10.0

20.0

30.0

40.0

50.0

60.0

70.0

80.0

1991

1992

1993

1994

1995

1996

1997

1998

1999

2000

2001

2002

2003

2004

2005

2006

2007

2008

2009

2010

2011

2012

2013

2014

Diesel Share Over Time

SPAIN Weighted Average

Notes: Authors’ calculations. New passenger car registration data from Association Auxiliaire de l’Automobile(AAA). European diesel penetration (dashed line) constructed gross domestic product as weights (source: WorldBank Development Indicators). Countries included: Austria, Belgium, Denmark, France, Germany, Ireland, Italy,Luxembourg, Netherlands, Portugal, and the United Kingdom.

– v –

B Estimation: Solving the Model

In this section we describe our algorithm to solve the model conditional on parameter guess θ =

[β,Σ,Π, γ]. Since solving the model is independent across years, we drop the t subscripts for brevity.

The algorithm is as follows:

1. Compute δ using the contraction mapping described in (Berry et al., 1995, Appendix I).

2. Use µ(Σ,Π) and sijt(θ) to solve for the implied markups bj . Use the firms’ first-order

conditions for the pricing game (Equation 7) and the observed prices to construct marginal

costs (c).

3. Use δ from (1) and the β parameter vector guess to solve for ξ.

4. Use the marginal costs (c) from (2) and the γ parameter vector guess to solve for ω.

5. Use the assumed information structure Ψ to construct the structural error νkj :

νkj (θ; Ψ) = sj(θ; Ψ)×∂(pτj (θ; Ψ)− cj(θ; Ψ))

∂xkj+∑r∈Ff

(pτr (θ; Ψ)− cr(θ; Ψ))× ∂sr(θ; Ψ)

∂xkj. (B.1)

using the following algorithm:

(a) Use γ and the Cobb-Douglas specification of the marginal cost equation to generate∂cj∂Xk

j

.

(b) Evaluate the indirect (price-induced) market share response to attributes from:

∂sr∂xkj

=

∫νk

∫D

(βk + σkνk + πkD)× sij(1− sir)dPD(D)dPν(ν) +∑m∈Ff

∂sr∂pm

∂pm∂xkj

, r = j,

−∫νk

∫D

(βk + σkνk + πkD)× sijsirdPD(D)dPν(ν) +∑m∈Ff

∂sr∂pm

∂pm∂xkj

, otherwise .

(B.2)

where we solve for pτm∂xkj

using the market shares and implicit function theorem. Note that

the dependence upon Ψ in Equation B.2 is assumed (and therefore dropped from the

notation for brevity). Equation B.2 requires solving for a fixed point in the matrix ∂s∂x .

While we have no proof that our operator is a contraction or that it results in a unique

solution, we found that updating guesses using a convex combination of the previous and

new guess yielded fast, monotonic convergence and that starting from different initial

guesses yielded the same results.

(c) Since we allow for Ψ = Ψft to vary by firm (i.e., firms may have different beliefs about

the future), we solve for the structural errors (ν) firm-by-firm.

– vi –

C EPA Cost Estimates for Retrofitting Diesel Vehicles

The following information was taken from ‘Diesel Retrofit Devices.” Environmental Protection

Agency (http://www.epa.gov/cleandiesel/technologies/retrofits.htm), last updated Jan-

uary 23, 2013. As described in the text, the retrofitting technology we consider is the “Lean NOx

Catalyst (LNC)” as this technology is most relevant for limiting NOx emissions in passenger cars.

Our inclusion of the remaining technologies recommended by the EPA shows both the breadth of

technologies available to reduce a variety of emissions as well as the variety of costs (of which the

LNC is near the bottom) required to retrofit a vehicle.

Diesel retrofit devices for after-treatment pollution control can be installed on new or

existing vehicles and equipment to reduce particulate matter (PM), nitrogen oxides

(NOx), hydrocarbons (HC), or carbon monoxide (CO) as well as other air pollutants.

The information below provides estimated emission reductions.

Table C.1: Estimated Costs to Retrofit Diesel Vehicles

Typical NOxTechnology Emission Reduction Typical Cost ($)

Lean NOx Catalyst (LNC) 5-40% $6,500-$10,000Selective Catalytic Reduction (SCR) <75% $10,000-$20,000; Urea $0.80/ gallon

Source: United States Environmental Protection Agency.

Lean NOx Catalyst (LNC) Lean NOx Catalysts (LNC) use diesel fuel injected into the

exhaust stream to create a catalytic reaction and reduce pollution. Verified LNCs are

paired with either a DPF or a DOC. An LNC can also be paired with an active DPF

to reduce NOx emissions and enable filter regeneration over a range of duty cycles.

However, an LNC still requires specific exhaust temperatures for appropriate NOx

emission control performance. LNCs can increase fuel usage by 5-7 percent (emphasis

added).

Selective Catalytic Reduction (SCR) Selective Catalytic Reduction (SCR) Systems in-

ject a reductant, also known as diesel exhaust fluid (DEF), into the exhaust stream

where it reacts with a catalyst to convert NOx emissions to N2 (nitrogen gas) and

oxygen. The catalytic reaction requires certain temperature criteria for NOx reduction

to occur. As with DPFs, knowing the age and type of each engine in the fleet as well

as the drive cycles of the vehicles is important. Data logging must be performed to

determine if the exhaust gas temperatures meet the specific SCR system requirements.

SCR systems require periodic refilling of the DEF, and the system should ensure that

the DEF never freezes. SCR systems are commonly used in conjunction with a DOC

and/or DPF to reduce PM emissions. Because of new NOx standards, most 2010 and

– vii –

newer on-highway diesel engines come equipped with an SCR system. A DEF refueling

infrastructure is in place, facilitating the use of SCRs.

D Solving for Counterfactual Automobile Prices

In this section we provide computational details to find the profit-maximizing prices under each

policy experiment. For the sake of brevity, we suppress the period subscripts. Each firm f produces

some subset Ff of the j = 1, . . . , J automobile brands and chooses a vector of pre-tariff prices pτj to solve:

maxpτj

∑j∈Ff

pτj − cj×Msj , (D.1)

The firm’s first-order condition for price conditional on product characteristics is given by:

sj +∑r∈Ff

(pτr − cr)×∂sr∂pτj

= 0 . (D.2)

Optimality requires that Equation (D.2) hold for all products sold in period t. We express the set

of firm f first-order conditions in matrix notation as:

s+ ∆× (pτ − c) = 0 , (D.3)

where an element of the matrix Ω is defined as:

Ωjr =

∂sj∂pτr

, if j, r ⊂ Ff ,

0 otherwise .

(D.4)

For a given vector of marginal costs c, we use (D.3) to find the fixed point to the system of

equations – a common practice in the literature dealing with this class of models. To our knowledge

there exists no proof of convergence or uniqueness for this contraction operator and fixed point.

Our experience (as is common) is that convergence is monotonic and proceeds quickly. Further,

starting from different starting values yields an identical result.

E Fuel Taxation as a Policy Tool to Promote Diesels

Following the European Fuel Taxation Directive of the 1970s, diesel fuel received a favorable

treatment that has convinced many to conclude that the success of diesel vehicles in Europe was

due primarily to this favorable treatment of diesel fuel taxation. We argued in Section 2.2 that

the reduced diesel fuel tax rate was instrumental for the development of a diesel market niche that

eased the adoption of tdi (and likely influenced its development) and other improved diesel vehicles

– viii –

in the 1990s, two decades after the European Fuel Tax Directive was adopted. Given this initial

condition, it is unclear to what degree preferential fuel taxes influenced the domestic market versus

the diesel-friendly environmental emissions policy. Moreover, both policies have the potential to

protect domestic industry by promoting its competitive advantage among consumers. The goal of

this appendix is to quantify the impacts of each to assess the relative impact on the industry.

In our first experiment (third panel of Table E.1) we replace the European fuel taxes

with average values employed in the United States where taxes are not only lower but also favor

gasoline.32 The reduction in fuel price increases their fuel efficiency, kpe, and consequently the

attractiveness of new cars (since σkpe > 0), increasing total consumption 10.6 percent, though the

increase is across both gasoline and diesel. The increase in diesel is more muted (3.7%) than gasoline

(18.7%) and both European firms and Non-European firms experience significant increases in profits

(9.7% and 11.6%, respectively). Conversely, quantity sold for diesels in our estimated equilibrium

(first panel) are 3.5% lower than under U.S. fuel taxes and both European firms and Non-European

firms experience are worse off under the current tax policy (8.8% and 10.4%, respectively).

In the fourth and fifth panels we evaluate the consequences of equalizing fuel taxes and

increasing fuel taxes by 8.1% in line with current policy.33 When we increase fuel taxes to the level

applied to gasoline, the higher fuel prices and lower fuel economy lead consumers to substitute

towards gasoline. Consequently, diesel sales and profits in the estimated equilibrium are 5.4%

and 5.4% greater in the estimated equilibrium. We see similar results in the current EU fuel

taxation policy where higher diesel fuel taxes lead consumers to substitute away from diesel varieties

indicating that the policy employed in the 1990s did the opposite – it encouraged consumers to

purchase the diesel cars largely produced by domestic auto makers.

We compare these results to the market equilibrium when auto makers are required to meet

stricter emissions requirements on NOx emissions (panel 2). Here, we use the lower-bound on the

EPA retrofitting cost as a conservative estimate. While pro-diesel fuel taxes increased consumption

of diesels around six percent, the pro-diesel emissions policy employed by the EU increased total

sales of diesels by 61.1% and most of these gains were captured by European automakers – profits

for EU automakers in our estimated equilibrium increased e610 million (21.9%). Total profits for

non-European auto makers also increases since some of these firms had adopted diesels though the

results are meager compared to their European rivals. These results indicate that while preferential

fuel taxes did play a role in promoting diesels and protecting domestic auto makers, fuel taxes play

a minor role compared to the diesel-friendly emissions policy employed by EU regulators.

32 In constructing these average tax rates we computed the average fuel taxes across states, weighting by aggregatestate fuel usage.

33Finally, after almost two decades of deliberation and negotiation among European policy makers, the EuropeanFuel Tax Directive of the 1970s was updated to account for the energy content of each type of fuel (instead of justits volume) as well as for their disparate environmental impact. These new taxation principles were supposed toeliminate the favorable taxation of diesel fuels among others. Excise fuel taxes at the bottom panel of Table E.1 arethose in place during 2015 according to E.U. Technical Press Briefing available at http://ec.europa.eu/taxation_customs/resources/documents/taxation/review_of_regulation_en.pdf

– ix –

Table E.1: Modifying Diesel Fuel Taxes

Benchmark Diesel and Gas Excise Taxes

Fuel Tax Price Quantity Margin Share Profit

eu: diesel 0.23 16.19 695.37 18.68 50.95 1,961.00eu: gasoline 0.35 14.93 508.70 21.09 37.28 1,434.37non-eu: diesel 0.23 17.20 36.97 14.84 2.71 83.26non-eu: gasoline 0.35 13.66 123.65 21.18 9.06 291.05

Total 0.29 15.52 1,364.70 19.70 100.00 3,769.68

Retrofitting Expense of $3,600 Euros

Fuel Tax Price Quantity Margin Share Profit

eu: diesel 0.23 20.02 462.89 15.12 40.15 1,336.27eu: gasoline 0.35 14.90 537.88 21.08 46.66 1,514.13non-eu: diesel 0.23 21.24 21.40 12.23 1.86 49.52non-eu: gasoline 0.35 13.65 130.64 21.18 11.33 307.27

Total 0.30 16.93 1,152.81 18.53 100.00 3,207.18

US Fuel Taxes

Fuel Tax Price Quantity Margin Share Profit

eu: diesel 0.15 16.11 721.96 18.73 47.82 2,033.11eu: gasoline 0.14 14.52 606.24 21.31 40.16 1,691.01non-eu: diesel 0.15 17.14 37.16 14.84 2.46 83.47non-eu: gasoline 0.14 13.29 144.35 21.35 9.56 334.35

Total 0.14 15.23 1,509.71 19.92 100.00 4,141.95

Diesel and Gas Excise Taxes are the Same

Fuel Tax Price Quantity Margin Share Profit

eu: diesel 0.35 16.24 658.35 18.62 49.34 1,855.48eu: gasoline 0.35 14.92 514.78 21.10 38.58 1,451.84non-eu: diesel 0.35 17.25 36.20 14.82 2.71 81.63non-eu: gasoline 0.35 13.66 125.02 21.19 9.37 294.38

Total 0.35 15.52 1,334.34 19.71 100.00 3,683.32

Diesel Excise Tax is Increased

Fuel Tax Price Quantity Margin Share Profit

eu: diesel 0.38 16.25 651.51 18.61 49.03 1,835.82eu: gasoline 0.35 14.92 515.92 21.10 38.83 1,455.11non-eu: diesel 0.38 17.26 36.04 14.82 2.71 81.30non-eu: gasoline 0.35 13.66 125.27 21.19 9.43 295.00

Total 0.37 15.52 1,328.74 19.72 100.00 3,667.22

Notes: Results based on year 2000 equilibrium. “Fuel Tax” is measured in 1994 Euros per liter and Total is thesales-weighted average fuel excise tax. “Price” is the sales-weighted average price faced by consumers (in thousandsof 1994 Euros), including tariffs. “Quantity” is measured in thousands of cars. “Profit” is measured in millions of1994 Euro. “Margin” and “Share” are reported as percentages.

– x –

F Additional Results

F.1 Unobserved Characteristics In Figure F.1 we present the evolution of estimated unobserv-

able quality ξ by year and fuel type. Table F.1 reports tests of stochastic dominance for these

distributions. It is remarkable that while the unobservable attributes of gasoline vehicles are

indistinguishable at the beginning and end of the 1990s, the perceived quality of diesels clearly

improved during that same time period. Consumers were uncertain about unobservable features

such as durability, torque, or reliability at the introduction of tdi. We also find that not only

diesels (or consumers’ perception of diesels) improve during the 1990s but that they are also linked

to power, size, brand, and other observable automobile attributes.

Figure F.1: Change in the Distribution of Unobserved Attributes

(a) Gasoline: ξ (b) Diesel: ξ

Table F.1: Distribution of Attributes

2000 vs 1992 1992 vs 2000

SD1 SD2 SD1 SD2

gasolinec90 0.202 0.207 0.723 0.509kpe 0.000 0.000 1.000 0.789size 0.697 0.825 0.454 0.273hpw 1.000 0.830 0.024 0.003

ξ 0.798 0.532 0.202 0.174

dieselc90 0.845 0.670 0.000 0.000kpe 0.000 0.000 1.000 0.780size 1.000 0.865 0.000 0.000hpw 0.002 0.123 0.000 0.000

ξ 1.000 0.736 0.000 0.000

Kolmogorov-Smirnov tests of first (SD1) and second (SD2) orderstochastic dominance where reported p-values are based on the con-sistent inference of Barrett and Donald (2003) using 1000 replicationsand 100 grid points on two random samples, for 1992 and 2000, of athousand draws from the kernel distribution densities of each attribute.A p-value smaller than 0.05 rejects the null stochastic dominancehypothesis.

– xi –

F.2 Substitution Patterns In Figure F.2 we present evidence that our model generates reasonable

substitution patterns. We show this by first solving for the distance between each pair of products

in a particular characteristic (e.g., hpw). We then divide the product-pairs into deciles where the

first decile correspond to pairs which are most alike. Finally, we compute the average cross-price

elasticity for each bin. The results are plotted in panels (a-c) where we see clearly that for all

of the characteristics considered in our estimation, substitution between similar products is much

more likely than for products far apart in characteristic space. Since diesel is a discrete variable,

we show the average cross-price elasticity within and across fuel types (panel d). Again, we see

that consumers are much more likely to substitute within fuel type

Figure F.2: Cross-Price Elasticities

0

.005

.01

.015

.02

.025

.03

Cro

ss-P

rice

Ela

stic

ity

0 10 20 30 40 50 60 70 80 90

Distance in Product Space (0 = Near, 90 = Far)

(a) hpw

0

.005

.01

.015

.02

.025

.03

Cro

ss-P

rice

Ela

stic

ity

0 10 20 30 40 50 60 70 80 90

Distance in Product Space (0 = Near, 90 = Far)

(b) size

0

.005

.01

.015

.02

.025

.03

Cro

ss-P

rice

Ela

stic

ity

0 10 20 30 40 50 60 70 80 90

Distance in Product Space (0 = Near, 90 = Far)

(c) kpe

0

.02

.04

.06

.08

Cro

ss-P

rice

Ela

stic

ity

1992 1996 2000

Dies-Dies Gas-Gas Diesel-Gas

(d) Fuel Type

– xii –

F.3 Robustness In Table F.2 we show that although the analysis focused on the year 2000, our

conclusions extend across the 1990s.

Table F.2: Implicit Tariff by Year and Estimation Approach

Import Baseline No Trend BLP

Year Tariff LB UB LB UB LB UB

1992 14.40 16.29 17.23 16.29 17.23 16.84 18.481993 10.30 12.67 13.81 13.22 14.15 13.35 15.461994 10.30 12.55 13.83 13.69 14.60 12.95 14.931995 10.30 12.92 14.64 15.05 16.04 13.73 16.371996 10.30 12.78 14.54 15.76 16.83 13.67 16.081997 10.30 12.74 14.63 16.56 17.53 13.12 15.501998 10.30 12.97 15.27 18.73 19.87 13.43 16.041999 10.30 13.58 16.52 21.82 23.40 14.22 17.622000 10.30 13.41 16.36 23.41 24.78 14.09 17.44

Notes: “Import Tariff” is the official import tariff placed on foreign imports. Lowerbound (“LB”) and upper bound (“UB”) retrofitting estimates based on installinga Lean NOx Catalyst (LNC). Technical and cost details located in Appendix C.

– xiii –